Portrait of An Thai Le

Robotics · Learning · Control
Optimization · Geometry · Relativity

An Thai Le

Assistant Professor, VinUniversity
Visiting Professor, TU Darmstadt · Director of Foundation AI, VinRobotics

I study how robots can plan, learn, and act reliably with limited computation and data.

Our group develops parallel motion planners, data- and compute-efficient robot foundation models, and force-aware controllers for heavy humanoids. We study locomotion and manipulation where robots must reason about physical contact.

Interested in working together? I welcome curious students and collaborators.

Latest news

Earlier news

Research

Efficient algorithmic robotics

I study how the way we represent a problem changes what a robot can compute and learn. Can we turn a sequential search into parallel operations? Can a policy keep its useful behavior with less data and compute? I work on these questions in motion planning, robot foundation models, and humanoid control, where the answers have to hold up on real hardware.

01 / Problem representation GTMP represents motion planning as a multipartite graph, with sampled configurations arranged in layers and candidate paths connecting them
Expose parallel structure Change the representation of search so many candidates share the same computation. GTMP · multipartite search graph
02 / Robot foundation models The vla.simd paper figure maps linear layers, convolutions, and attention onto tiled matrix operations and a shared SIMD micro-kernel with a fixed register budget
Scale down, retain capability Use structure and shared computation to fit learned behavior into a smaller budget. vla.simd · native SIMD computation
03 / Whole-body control A simulated humanoid bends its knees under a downward force at the pelvis; a red arrow indicates the applied force
Reason through contact Coordinate motion and compliance across the body as forces and support change. CompliantWBC · force response

01 / Planning & optimization

Representations for efficient algorithms

How can a change of representation simplify an algorithm and expose more parallel computation?

Many planners explore one possibility at a time. By arranging candidate robot configurations in layers, we can evaluate many possible paths together on a GPU. I study how choices like this change both the algorithm and the computation it needs.

Tensor motion planning, optimal transport, and learned trajectory priors give us different ways to organize that search. I care about when they find a solution and whether more computation improves it. Our completeness and convergence results make the assumptions behind those answers explicit.

GTMP · Anytime GTMP · MPOT · MTP · MPD · CLOT · PolyStep · AAC

A simulated Panda arm follows a planned path around a box
GTMP · parallel motion planning. Layered search graph
Model Tensor Planning controls a simulated robot through repeated trajectory planning
MTP · tensor search for model predictive control.

02 / Robot foundation models

Scaling down robot foundation models

How much capability can we retain with less data, memory, and compute?

I want robot foundation models to be useful on the hardware we can put on a robot. We remove redundant layers and use low-bit arithmetic to reduce inference cost. To learn from fewer demonstrations, we build geometric symmetries into policies and use future observations as supervision.

A smaller model still needs an implementation that suits the hardware. Memory layout and shared operations can matter as much as parameter count. We also use search and retrieval to guide a frozen policy during execution. For each approach, we measure what the robot can still do alongside latency and memory use.

FoldQuantVLA · vla.simd · vla.cpp · CLP · FOCA · FINE · EquiVLA · PAINT · SDN · R2A

An SO-101 arm running IMPACT picks up a roll of tape and places it in a box in the vla.simd hardware demonstration
vla.simd / IMPACT · "Put the tape into the box." SO-101, original speed.
PAINT hardware experiments with towel flinging on ALOHA and part placement on a humanoid
PAINT · connecting action chunks without retraining. 2× speed.

03 / Humanoid control

Whole-body control through contact

How should a humanoid use its whole body to respond to contact and changing loads?

A pull on a humanoid's arm changes what its torso and legs must do to keep balance. I study controllers that respond through the whole body, using force estimates from the robot's own sensors, parallel foothold search, and compliance training. We test these ideas on 70–85 kg humanoids under contact and changing loads.

We also ask what a controller can reuse when the robot's body changes. CrossBFM transfers a shared behavior space, with the strongest transfer between similar humanoids. In γ₀, we train one motion policy across many robot bodies. Both approaches help us study which parts of behavior can be shared across hardware.

γ₀ · CrossBFM · CompliantWBC · TACT-ful · DoublyAware · Hardware experiments

CompliantWBC: a 70 kg humanoid bends its knees under a person's downward pull, then rises again
CompliantWBC · whole-body response to a downward pull.
Fourier N1 and Booster T1 humanoids perform related motions in a CrossBFM simulation
CrossBFM · shared behavior across humanoids.

The bigger picture

The geometry of computation

I keep coming back to the connection between a problem's geometry and the computation it needs. Nearby trajectories can share work during search. Symmetries can reduce what a policy has to learn. A robot's configuration determines how it can exert or withstand a force. I want our algorithms to make use of these relationships.

  1. Problem geometryTrajectories, symmetries, contact
  2. Computational structureBatches, shared work, coordination
  3. Hardware organizationParallel lanes, memory, native kernels

By the geometry of computation, I mean how problem structure maps onto hardware. Tensor planning makes candidate paths regular enough to evaluate in batches; vla.simd organizes shared work around CPU instructions; foothold search brings parallel candidate evaluation into locomotion.

My longer-term goal is to bring these pieces together: policies and planners that fit the robot's compute budget, connected to controllers that regulate motion and compliance. We want to test the combined system on humanoid tasks involving contact, measuring task success alongside data requirements, memory use, and response time.

A personal interest

Computational general relativity

An Alcubierre-bubble parameter sweep shows expansion, Eulerian energy density, and the null-energy-condition margin
warpax · expansion and energy-condition diagnostics for an Alcubierre-bubble parameter sweep.

I write numerical tools for general relativity in my free time because, to me, the physics is art. warpax computes spacetime curvature with vectorized auto-diff in JAX and checks energy conditions for all observers at a point through matrix inequalities. These pointwise tests alone cannot establish whether a warp drive is physically realizable. The Kerr-flyby study simulates rocket trajectories to find when energy extraction from a rotating black hole still allows escape. I use the same design thinking as in robotics: batch the computation and make the assumptions explicit.

warpax · Classical and Quantum Gravity · Penrose extraction · Physical Review D

In the news: warpax in Popular Mechanics (May 2026); VinRobotics' visit to DFKI; VinRobotics humanoids at ICRA 2026 in Báo Tin Tức.

Humanoids at work

Hardware experiments with our VinRobotics team: handling parts, performing martial arts, responding to contact, and walking under load.

A humanoid picks and places parts between trays at a kitting workstation

Kitting

Part-handling pilot.

A humanoid balances on one leg while performing a high kick

Humanoid high kick

Whole-body motion imitation on a humanoid.

A humanoid yields to external contact through its arms, torso, and legs while balancing

Compliant whole-body control

Coordinating the whole body to respond to contact.

A humanoid carrying a 20 kg backpack descends laboratory stairs

Stair descent with a 20 kg payload

TACT-ful combines foothold planning with compliance training.

More humanoid videos · Research behind the demos

Publications

Papers and ongoing projects, each with its current status. * marks co-first or co-last authors. A blue line marks representative papers. Select a figure to view it at full size. See also Google Scholar.

Topic
Year

2026

  1. γ₀: A Generalist Policy for Multi-Embodiment Motion Control

    Nico Bohlinger and Bo Ai (co-leads), with the γ₀ team · An Thai Le (advisor)

    We train a single motion policy across varied robot bodies, randomizing their embodiments during training to learn what carries over between them. The same policy controls humanoids and quadrupeds at the joint level, with outdoor hardware demonstrations. Try it on different robots in the interactive playground.

    Ongoing project · 2026Project pagePlayground
    A Unitree G1 humanoid walks outdoors alongside Unitree Go2 and Silver Badger quadrupeds, all controlled by Gamma Zero

  2. CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments

    T.D. Do*, T.D. Phuong*, N. Bohlinger, C.T. Trinh, S. Ju, V.A. Ngo, J. Peters, X. Wang*, An Thai Le*

    Transfers a shared behavior space across humanoids. We use corresponding frames of retargeted motion to train an encoder without simulation, then train a controller for each robot. The shared space supports motion tracking, goal reaching, and reward prompts, with hardware demonstrations. Transfer to an unseen robot is strongest when its body resembles those used in training.

    Under review · 2026arXivProject page
    BibTeX
    @misc{do2026crossbfm,
  title         = {{CrossBFM}: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments},
  author        = {Do, Tan-Dzung and Phuong, Tuan Dat and Bohlinger, Nico and Trinh, Cuc T. and Ju, Siwei and Ngo, Vien Anh and Peters, Jan and Wang, Xinchao and Le, An Thai},
  year          = {2026},
  eprint        = {2609.38087},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.38087}
}

  3. CompliantWBC: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets

    T.D. Do*, C.T. Trinh*, T.D. Phuong, T.D. Dang, C. Le, T. Ly, V.A. Ngo, An Thai Le

    Estimates external forces from proprioception and adjusts impedance targets so a 70 kg humanoid yields to contact across its body. The analytical passivity guarantee is local and assumes fixed stiffness.

    Under review · 2026Project pagePDF
    BibTeX
    @misc{do2026compliantwbc,
  title        = {{CompliantWBC}: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets},
  author       = {Do, Tan-Dzung and Trinh, Cuc T. and Phuong, Tuan Dat and Dang, Truong-Duy and Le, Chien and Ly, Thanh and Ngo, Vien Anh and Le, An T.},
  year         = {2026},
  howpublished = {\url{https://compliantwbc.github.io/paper/compliant_wbc.pdf}},
  note         = {Under review}
}

  4. TACT-ful: Multi-Channel Terrain Affordance and Compliance Training for Payload-Robust Perceptive Humanoid Locomotion

    T. Ly*, T.D. Dang*, C. Le, T.D. Do, T.D. Phuong, C.T. Trinh, V.A. Ngo, An Thai Le

    Combines GPU-parallel foothold search with compliance training to help humanoids walk over uneven terrain with changing payloads. Evaluated quantitatively in simulation, with qualitative hardware demonstrations.

    Under review · 2026arXivProject page
    BibTeX
    @misc{ly2026tactful,
  title         = {{TACT}-ful: Multi-Channel Terrain Affordance and Compliance Training for Payload-Robust Perceptive Humanoid Locomotion},
  author        = {Ly, Thanh and Dang, Truong-Duy and Le, Chien and Do, Tan-Dzung and Phuong, Tuan Dat and Trinh, Cuc T. and Ngo, Vien Anh and Le, An T.},
  year          = {2026},
  eprint        = {2606.20645},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.20645}
}

  5. Pretraining Foundation Policies for Perceptive Humanoid Locomotion with Offline RL SOFIE

    Learns humanoid locomotion offline from expert rollouts of mixed quality. We score whole footsteps and give better examples more weight, so the policy is less likely to copy the experts' mistakes. A shared policy walks three humanoids over stepping stones, stairs, and rough ground. Simulation tests cover twelve humanoids across two simulators; pretraining also helps fine-tuning on an unseen humanoid.

    Under review · 2026Project page
    BibTeX
    @misc{sofie2026,
  title        = {Pretraining Foundation Policies for Perceptive Humanoid Locomotion with Offline {RL}},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {\url{https://silentgradient.github.io/SOFIE/}},
  note         = {Under review}
}

  6. Terrain Traversal Without Motion Priors: Expert-Routed Humanoid Locomotion at Deployment Fidelity

    Learns perceptive humanoid locomotion without motion-capture retargeting, a generative motion prior, or distillation. A terrain classifier selects among five experts, and gradient steps refine each foothold using a terrain-cost map. The policy completes an 85 m course in 18 of 20 runs in a second simulator configured for deployment. Hardware tests are still to come.

    Under review · 2026Project page
    BibTeX
    @misc{moeterrain2026,
  title        = {Terrain Traversal Without Motion Priors: Expert-Routed Humanoid Locomotion at Deployment Fidelity},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {\url{https://fai-rl-tech.github.io/moe-perceptive/}},
  note         = {Under review}
}

  7. DoublyAware: Dual Planning and Policy Awareness for Temporal Difference Learning in Humanoid Locomotion

    Khang Nguyen, An Thai Le, Jan Peters, Minh Nhat Vu

    Handles planning and policy uncertainty separately in TD-MPC: conformal filtering screens candidate trajectories, and a group-relative trust region constrains policy updates. Evaluated on a simulated H1-2 humanoid.

    IEEE RA-L 2026DOIarXiv
    BibTeX
    @article{nguyen2026doublyaware,
  title         = {{DoublyAware}: Dual Planning and Policy Awareness for Temporal Difference Learning in Humanoid Locomotion},
  author        = {Nguyen, Khang and Le, An Thai and Peters, Jan and Vu, Minh Nhat},
  journal       = {IEEE Robotics and Automation Letters},
  volume        = {11},
  number        = {2},
  pages         = {2162--2169},
  year          = {2026},
  doi           = {10.1109/LRA.2025.3648611},
  eprint        = {2506.12095},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  8. FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding

    Hung T. Ho, Khanh D. Nguyen, Quang D. Nguyen, Thanh Q. Duong, Ngan Le, Meng Guo, Vien A. Ngo, An Thai Le

    Quantizes VLA policies for native low-bit inference without retraining, using a consistent activation representation. W4A4 projections run 1.20–1.33× faster than floating-point TensorRT on Jetson AGX Orin; selective INT8 improves action fidelity and measured real-robot success.

    Preprint · 2026arXivProject pageCode
    BibTeX
    @misc{ho2026foldquantvla,
  title         = {FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding},
  author        = {Ho, Hung T. and Nguyen, Khanh D. and Nguyen, Quang D. and Duong, Thanh Q. and Le, Ngan and Guo, Meng and Ngo, Vien A. and Le, An T.},
  year          = {2026},
  eprint        = {2609.24433},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.24433}
}
    FoldQuantVLA: consistent folding and native low-bit policy execution

  9. vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation

    Khanh D. Nguyen, Hoang M. Truong, An Thai Le

    Runs VLA policies on CPUs by sharing SIMD kernels and reusing computation, with demonstrations on SO-101 and UR10e. On Raspberry Pi 5, IMPACT outputs 33.5 actions/s in fp32 after a 90 s thermal soak. This measures action throughput, not feedback frequency.

    Preprint · 2026arXivProject pageCode
    BibTeX
    @misc{nguyen2026vlasimd,
  title         = {vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation},
  author        = {Nguyen, Khanh D. and Truong, Hoang M. and Le, An T.},
  year          = {2026},
  eprint        = {2609.24274},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.24274}
}
    vla.simd: shared SIMD micro-kernels for policy inference on CPUs

  10. Training Non-Differentiable Networks via Optimal Transport PolyStep

    An Thai Le

    Uses optimal transport to weight candidate parameter updates and train non-differentiable networks with only forward evaluations. Compares gradient-free methods at matched evaluation budgets.

    TMLR 2026OpenReviewarXivCode
    BibTeX
    @article{le2026training,
  title         = {Training Non-Differentiable Networks via Optimal Transport},
  author        = {Le, An T.},
  journal       = {Transactions on Machine Learning Research},
  issn          = {2835-8856},
  year          = {2026},
  url           = {https://openreview.net/forum?id=8mlcqTTMuU},
  eprint        = {2605.01928},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG}
}

  11. Observer-robust energy condition verification for warp drive spacetimes warpax

    An Thai Le

    Uses vectorized auto-diff in JAX to check energy conditions in warp drive spacetimes. Combines 4×4 matrix inequalities over all observers at a point with interval bounds for certified decisions.

    Class. Quantum Grav. 2026DOIarXivCode
    BibTeX
    @article{le2026observer,
  title         = {Observer-robust energy condition verification for warp drive spacetimes},
  author        = {Le, An T.},
  journal       = {Classical and Quantum Gravity},
  year          = {2026},
  month         = oct,
  doi           = {10.1088/1361-6382/aeb25a},
  eprint        = {2602.18023},
  archivePrefix = {arXiv},
  primaryClass  = {gr-qc}
}

  12. Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection PAINT

    T.B. Ho, Q.T. Nguyen, T.L. Ha, G.B. Nguyen, V.T. Nguyen, L. Dinh, M.N. Vu, D.M.H. Nguyen, An Thai Le, V.A. Ngo

    Uses flow inversion to select initial noise that smoothly connects action chunks during asynchronous execution. Requires no retraining or inference-time gradients; evaluated on 12 simulated benchmarks and 6 real manipulation tasks.

    CoRL 2026arXivProject pageRepository
    BibTeX
    @inproceedings{ho2026start,
  title         = {Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection},
  author        = {Ho, Trong-Bao and Nguyen, Quang-Tan and Ha, Thien-Loc and Nguyen, Gia-Binh and Nguyen, Viet-Thanh and Dinh, Long and Vu, Minh N. and Nguyen, Duy M. H. and Le, An Thai and Ngo, Vien Anh},
  booktitle     = {Conference on Robot Learning (CoRL)},
  year          = {2026},
  eprint        = {2606.19774},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.19774}
}

  13. FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation

    D.M. Nguyen*, N.T. Diep*, G.B. Nguyen*, T.B. Ho, D.T. Le, Q.T. Nguyen, T.L. Ha, V.N. Tran, B. Thach, X.N. Tran, T.A. Tran, A. Habuda, P.L. Møller, T.N. Le, D. Sonntag, M. Niepert, K.D. Doan, V.N. Duong, H. Ngo, M.N. Vu, D.M.H. Nguyen*, An Thai Le*, V.A. Ngo*

    Adapts VLA policies from few demonstrations by learning representations of future interactions and goals. Reaches 95.7% success on LIBERO with 20 demonstrations per task. Also evaluated in RoboCasa and on real robots.

    ICML 2026arXivProject pageCode
    BibTeX
    @inproceedings{nguyen2026foca,
  title         = {{FOCA}: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation},
  author        = {Nguyen, Duc Minh and Diep, Nghiem Tuong and Nguyen, Binh Gia and Ho, Trong-Bao and Le, Doanh and Nguyen, Tan Q. and Ha, Thien-Loc and Tran, Nhiem and Thach, Bao and Tran, Nhat X. and Tran, Tuan A. and Habuda, Artur and M{\o}ller, Philip Lund and Le, Tran Nguyen and Sonntag, Daniel and Niepert, Mathias and Doan, Khoa D. and Duong, Vu and Ngo, Hung and Vu, Minh N. and Nguyen, Duy M. H. and Le, An Thai and Ngo, Vien Anh},
  booktitle     = {International Conference on Machine Learning (ICML)},
  series        = {Proceedings of Machine Learning Research},
  volume        = {306},
  publisher     = {PMLR},
  year          = {2026},
  eprint        = {2606.20867},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.20867}
}

  14. CLOT: Multi-Robot Motion Planning via Collaborative Optimal Transport under Signal Temporal Logic Tasks

    Y. Zhang, Y. Zhang, An Thai Le, M. Guo

    Uses collaborative optimal transport to plan multi-robot trajectories for tasks specified in signal temporal logic. Updates trajectories without gradients; evaluated in simulation and on hardware.

    ICRA 2026PDF
    BibTeX
    @inproceedings{zhang2026clot,
  title     = {{CLOT}: Multi-robot Motion Planning via Collaborative Optimal Transport under Signal Temporal Logic Tasks},
  author    = {Zhang, Ying and Zhang, Yunyi and Le, An T. and Guo, Meng},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
  url       = {https://mengguo.github.io/personal_site/papers/pdf/zhang2026clot.pdf}
}

  15. StructSAM: Structure- and Spectrum-Preserving Token Merging for Segment Anything Models

    D.M.H. Nguyen, T.A. Tran, D. Nguyen, S. Xie, T.Q. Nguyen, M.T.N. Truong, D. Palenicek, An Thai Le, M. Barz, E. Hannus, T. Nguyen, T. Dam, T.N. Le, N. Le, M. Vu, K. Doan, V. Ngo, P. Xie, J. Zou, D. Sonntag, J. Peters, M. Niepert

    Merges redundant tokens in Segment Anything's encoder while preserving boundaries and prompt regions. Measures the tradeoff between computation and segmentation accuracy.

    ICML 2026 Workshop (AdaptFM)arXivOpenReviewWorkshop
    BibTeX
    @inproceedings{nguyen2026structsam,
  title         = {{StructSAM}: Structure- and Spectrum-Preserving Token Merging for Segment Anything Models},
  author        = {Nguyen, Duy M. H. and Tran, Tuan A. and Nguyen, Duong and Xie, Siwei and Nguyen, Trung Q. and Truong, Mai T. N. and Palenicek, Daniel and Le, An T. and Barz, Michael and Hannus, Eric and Nguyen, TrungTin and Dam, Tuan and Le, Tran Nguyen and Le, Ngan and Vu, Minh and Doan, Khoa and Ngo, Vien and Xie, Pengtao and Zou, James and Sonntag, Daniel and Peters, Jan and Niepert, Mathias},
  booktitle     = {ICML 2026 Workshop on Resource-Adaptive Foundation Model Inference (AdaptFM)},
  year          = {2026},
  url           = {https://openreview.net/forum?id=K41rNB0F8j},
  eprint        = {2603.07307},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

  16. Rarity of rocket-driven Penrose extraction in Kerr spacetime Penrose rocket

    An Thai Le

    Simulates 320,000 rocket flybys around rotating black holes to test energy extraction and escape. Successful cases are rare in the broad scans and concentrated at high black-hole spin and relativistic exhaust speeds.

    Phys. Rev. D 2026DOIarXivCode
    BibTeX
    @article{le2026rarity,
  title         = {Rarity of rocket-driven {Penrose} extraction in {Kerr} spacetime},
  author        = {Le, An T.},
  journal       = {Physical Review D},
  volume        = {113},
  number        = {12},
  pages         = {124036},
  year          = {2026},
  month         = jun,
  doi           = {10.1103/f3jn-7wcp},
  eprint        = {2601.19616},
  archivePrefix = {arXiv},
  primaryClass  = {astro-ph.HE}
}

  17. Anytime Global Tensor Motion Planning Anytime GTMP

    S. Coumar, An Thai Le*, Z. Kingston*

    Extends tensor motion planning to black-box local planners, finding more paths and lowering their cost as computation continues. Asymptotic optimality holds under stated clearance and sampling assumptions.

    Under review · 2026arXivCode
    BibTeX
    @misc{coumar2026anytime,
  title         = {Anytime Global Tensor Motion Planning},
  author        = {Coumar, Sai and Le, An T. and Kingston, Zachary},
  year          = {2026},
  eprint        = {2608.25830},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.25830}
}

  18. TD-MTP: Coupling Structured Planning with Multi-Step Value Learning

    Asks whether a planner and its critic improve when changed together. Across ten paired tasks, adding a tensor-graph planner has different effects depending on whether the critic uses a one-step or TD(λ) target. On humanoid stairs, each change alone lowers return, but combining them improves on the baseline. We find clear evidence of this interaction on two of the ten tasks.

    Under review · 2026Project page
    BibTeX
    @misc{tdmtp2026,
  title        = {{TD-MTP}: Coupling Structured Planning with Multi-Step Value Learning},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {\url{https://tdmtp.github.io/}},
  note         = {Under review}
}

  19. EquiVLA: A General Framework for Rotationally Equivariant Vision-Language-Action Models

    T.L. Ha*, Q.T. Nguyen*, T.B. Ho*, L. Dinh, M.D. Nguyen, G.B. Nguyen, T.Q. Pham, M.N. Vu, D.M.H. Nguyen, An Thai Le, V.A. Ngo

    Adds rotational equivariance to pretrained VLAs so actions rotate with the input. The perception module is approximately equivariant; the action head is exactly equivariant. Evaluated with GR00T N1.5 on LIBERO.

    Under review · 2026arXivProject page
    BibTeX
    @misc{ha2026equivla,
  title         = {{EquiVLA}: A General Framework for Rotationally Equivariant Vision-Language-Action Models},
  author        = {Ha, Thien-Loc and Nguyen, Quang-Tan and Ho, Trong-Bao and Dinh, Long and Nguyen, Minh Duc and Nguyen, Gia-Binh and Pham, Tri Quang and Vu, Minh N. and Nguyen, Duy M. H. and Le, An Thai and Ngo, Vien Anh},
  year          = {2026},
  eprint        = {2606.19784},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.19784}
}

  20. vla.cpp: A Unified Inference Runtime for Vision-Language-Action Models

    K.D. Nguyen, H.T. Ho, C.T. Nguyen, T.Q. Duong, L.D. Le, D.M.H. Nguyen, V.A. Ngo, An Thai Le

    Runs eleven VLA models in a C++ runtime for onboard inference. BitVLA completes 200/200 simulated LIBERO-Object episodes with inference on an 8 GB Jetson Orin Nano in the reported evaluation.

    Under review · 2026arXivProject pageCode
    BibTeX
    @misc{nguyen2026vlacpp,
  title         = {{vla.cpp}: A Unified Inference Runtime for Vision-Language-Action Models},
  author        = {Nguyen, Khanh D. and Ho, Hung T. and Nguyen, Chinh T. and Duong, Thanh Q. and Le, Linh D. and Nguyen, Duy M. H. and Ngo, Vien A. and Le, An T.},
  year          = {2026},
  eprint        = {2606.08094},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.08094}
}

  21. Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think CLP

    G.B. Nguyen, T.B. Ho, T.L. Ha, K. Vo, P.L. Møller, Q.T. Nguyen, L. Dinh, T.M. Luu, T. Dam, V. Duong, T. Le, N.D.Q. Bui, M. Vu, T.N. Le, An Thai Le, N. Le, D. Sonntag, J. Zou, J. Peters, D.M.H. Nguyen, V.A. Ngo

    Prunes redundant VLA layers before fine-tuning with one pass that measures representation similarity. Evaluates how reducing model depth and training time affects task performance.

    Under review · 2026arXivProject pageCode
    BibTeX
    @misc{nguyen2026finetuning,
  title         = {Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think},
  author        = {Nguyen, Gia-Binh and Ho, Trong-Bao and Ha, Thien-Loc and Vo, Khoa and M{\o}ller, Philip Lund and Nguyen, Quang T. and Dinh, Long and Luu, Tung M. and Dam, Tuan and Duong, Vu and Le, Trung and Bui, Nghi D. Q. and Vu, Minh and Le, Tran Nguyen and Le, An Thai and Le, Ngan and Sonntag, Daniel and Zou, James and Peters, Jan and Nguyen, Duy M. H. and Ngo, Vien Anh},
  year          = {2026},
  eprint        = {2606.20246},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.20246}
}

  22. Self-Improving VLA Policies: Selected Diffusion Noise for Spurious-Robust Action Smoothing SDN

    D.M. Nguyen, B.N. Dao, T.M. Luu, B.G. Nguyen, V. Tong, A. Liu, V.N. Duong, D.D. Le, D. Sonntag, T. Le, N. Le, J. Peters, An Thai Le, M.N. Vu, M. Niepert, K.D. Doan, D.M.H. Nguyen, V.A. Ngo

    Selects initial diffusion noise for a frozen VLA by filtering for object-grounded action candidates, then executing the smoothest. Requires no policy retraining; evaluated in simulation and on real robots.

    Under review · 2026arXiv
    BibTeX
    @misc{nguyen2026self,
  title         = {Self-Improving {VLA} Policies: Selected Diffusion Noise for Spurious-Robust Action Smoothing},
  author        = {Nguyen, Duc Minh and Dao, Bao-Ngoc and Luu, Tung M. and Nguyen, Binh Gia and Tong, Vinh and Liu, Anji and Duong, Vu N. and Le, Dung D. and Sonntag, Daniel and Le, Trung and Le, Ngan and Peters, Jan and Le, An Thai and Vu, Minh Nhat and Niepert, Mathias and Doan, Khoa D. and Nguyen, Duy M. H. and Ngo, Vien Anh},
  year          = {2026},
  eprint        = {2606.14084},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.14084}
}

  23. RoboGaze: Evaluating Robot World Models via Structured Vision-Language Analysis

    M.L. Nguyen, N.T. Diep, H.K. Nguyen, M. Le, D.T. Le, H.H. Tran, D.D. Le, V.N. Duong, D. Sonntag, An Thai Le, D.M.H. Nguyen, V.A. Ngo, V.N. Tran

    Uses structured vision-language analysis to classify errors and identify when they occur in videos generated by robot world models, without training an evaluator.

    Under review · 2026arXivProject pageCode
    BibTeX
    @misc{nguyen2026robogaze,
  title         = {{RoboGaze}: Evaluating Robot World Models via Structured Vision-Language Analysis},
  author        = {Nguyen, Minh-Loi and Diep, Nghiem Tuong and Nguyen, Hung Khang and Le, Minh and Le, Doanh Thien and Tran, Hoang H. and Le, Dung D. and Duong, Vu N. and Sonntag, Daniel and Le, An Thai and Nguyen, Duy Minh Ho and Ngo, Vien Anh and Tran, Van Nhiem},
  year          = {2026},
  eprint        = {2606.28385},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.28385}
}

  24. FINE: Future-Informed Navigation Encoding for Data-Efficient Vision-Language Navigation

    K.H. Nguyen*, H.P.Q. Nguyen*, H.P. Nguyen, K.D. Binh, X.H. Nguyen, V.A. Ngo, D.M.H. Nguyen, H. Nguyen, An Thai Le

    Learns from future observations in navigation demonstrations. Landmark tokens predict upcoming landmarks' appearance and 3D geometry; a future token distinguishes the reached state from world-model counterfactuals. Gains on R2R-CE and RxR-CE grow with fewer demonstrations. This supervision adds no computation at deployment.

    Preprint · 2026arXivProject page
    BibTeX
    @misc{nguyen2026fine,
  title         = {{FINE}: Future-Informed Navigation Encoding for Data-Efficient Vision-Language Navigation},
  author        = {Nguyen, Khang H. and Nguyen, Hoang Pham Quang and Nguyen, Ha Phuong and Binh, Khanh Dinh and Nguyen, Xuan Ha and Ngo, Vien Anh and Nguyen, Duy M. H. and Nguyen, Huan and Le, An Thai},
  year          = {2026},
  eprint        = {2609.32855},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.32855}
}

  25. Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution R2A

    Builds a motion primitive graph from successful demonstrations. At test time, a query-conditioned graph network retrieves reference motion and blends it into a frozen VLA's actions through a per-dimension gate. Improves GR00T N1.7, π0.5, and StableVLA under camera, lighting, noise, and initial-state shifts in simulation and on a real Aloha robot, without fine-tuning.

    Under review · 2026Project page
    BibTeX
    @misc{r2a2026,
  title        = {Retrieve to Act: Motion Primitive Graph Retrieval for Robust {VLA} Execution},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {\url{https://retrieve2act.github.io/r2a/}},
  note         = {Under review}
}

  26. AAC: Admissible-by-Architecture Differentiable Landmark Compression for ALT

    An Thai Le, V.A. Ngo

    Learns to compress landmark heuristics for ALT graph search while preserving admissibility: the heuristic never overestimates the remaining path cost. Evaluated on road networks against fixed landmark-selection baselines.

    Under review · 2026arXivCode
    BibTeX
    @misc{le2026aac,
  title         = {{AAC}: Admissible-by-Architecture Differentiable Landmark Compression for {ALT}},
  author        = {Le, An T. and Ngo, Vien},
  year          = {2026},
  eprint        = {2604.20744},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2604.20744}
}

2025

  1. Model Tensor Planning MTP

    An Thai Le, Khai Nguyen, Minh Nhat Vu, João Carvalho, Jan Peters

    Combines global candidates from layered graphs with local trajectory sampling for model predictive control. Batches simulator rollouts in MuJoCo XLA; available in hydrax.

    TMLR 2025, ICLR 2026 (J2C)OpenReviewCode
    BibTeX
    @article{le2025model,
  title         = {Model Tensor Planning},
  author        = {Le, An Thai and Nguyen, Khai and Vu, Minh Nhat and Carvalho, Jo{\~a}o and Peters, Jan},
  journal       = {Transactions on Machine Learning Research},
  issn          = {2835-8856},
  year          = {2025},
  url           = {https://openreview.net/forum?id=fk1ZZdXCE3},
  eprint        = {2505.01059},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  2. Global Tensor Motion Planning GTMP

    An Thai Le, Kay Pompetzki, João Carvalho, Joe Watson, Julen Urain, Armin Biess, Georgia Chalvatzaki, Jan Peters

    Searches a layered graph with tensor operations to produce batches of smooth spline paths in parallel. Provides probabilistic completeness under the paper's assumptions.

    IEEE RA-L 2025, ICRA 2026DOIarXivCode
    BibTeX
    @article{le2025global,
  title         = {Global Tensor Motion Planning},
  author        = {Le, An Thai and Pompetzki, Kay and Carvalho, Jo{\~a}o and Watson, Joe and Urain, Julen and Biess, Armin and Chalvatzaki, Georgia and Peters, Jan},
  journal       = {IEEE Robotics and Automation Letters},
  volume        = {10},
  number        = {7},
  pages         = {7302--7309},
  year          = {2025},
  doi           = {10.1109/LRA.2025.3575307},
  eprint        = {2411.19393},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  3. Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models MPD

    João Carvalho, An Thai Le, Piotr Kicki, Dorothea Koert, Jan Peters

    Learns a diffusion prior over B-spline trajectories and guides sampling with task costs. Evaluated on 2D tasks and 7-DoF arms, including real pick-and-place tasks learned from human demonstrations.

    IEEE T-RO 2025, AAAI 2026 (journal track)DOIarXivCode
    BibTeX
    @article{carvalho2025motion,
  title         = {Motion Planning Diffusion: Learning and Adapting Robot Motion Planning With Diffusion Models},
  author        = {Carvalho, Jo{\~a}o and Le, An Thai and Kicki, Piotr and Koert, Dorothea and Peters, Jan},
  journal       = {IEEE Transactions on Robotics},
  volume        = {41},
  pages         = {4881--4901},
  year          = {2025},
  doi           = {10.1109/TRO.2025.3593109},
  eprint        = {2412.19948},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  4. FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching

    Khang Nguyen, An Thai Le, T. Pham, M. Huber, Jan Peters, Minh Nhat Vu

    Uses second-order conditional flow matching to learn acceleration-aware motion fields that generate smooth robot trajectories.

    IROS 2025arXivCode
    BibTeX
    @inproceedings{nguyen2025flowmp,
  title         = {{FlowMP}: Learning Motion Fields for Robot Planning with Conditional Flow Matching},
  author        = {Nguyen, Khang and Le, An Thai and Pham, Tien and Huber, Manfred and Peters, Jan and Vu, Minh Nhat},
  booktitle     = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages         = {11291--11297},
  year          = {2025},
  doi           = {10.1109/IROS60139.2025.11246537},
  eprint        = {2503.06135},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  5. Machine Learning with Physics Knowledge for Prediction: A Survey

    Joe Watson, Chen Song, Oliver Weeger, Theo Gruner, An Thai Le, Kay Hansel, Ahmed Hendawy, Oleg Arenz, Will Trojak, Miles Cranmer, Carlo D'Eramo, Fabian Bülow, Tanmay Goyal, Jan Peters, Martin W. Hoffmann

    Surveys how physical knowledge informs model architectures, training objectives, and data, focusing on partial differential equations and open-source tools.

    TMLR 2025arXiv
    BibTeX
    @article{watson2025machine,
  title         = {Machine Learning with Physics Knowledge for Prediction: A Survey},
  author        = {Watson, Joe and Song, Chen and Weeger, Oliver and Gruner, Theo and Le, An Thai and Hansel, Kay and Hendawy, Ahmed and Arenz, Oleg and Trojak, Will and Cranmer, Miles and D'Eramo, Carlo and B{\"u}low, Fabian and Goyal, Tanmay and Peters, Jan and Hoffmann, Martin W.},
  journal       = {Transactions on Machine Learning Research},
  issn          = {2835-8856},
  year          = {2025},
  url           = {https://openreview.net/forum?id=ZiJYahyXLU},
  note          = {Survey Certification},
  eprint        = {2408.09840},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG}
}

2024

  1. Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks ConAN

    D.M.H. Nguyen*, N. Lukashina*, T. Nguyen, An Thai Le, T. Nguyen, N. Ho, Jan Peters, D. Sonntag, V. Zaverkin, M. Niepert

    Predicts molecular properties by combining a 2D molecular graph with multiple 3D conformers using a differentiable fused Gromov–Wasserstein barycenter. Predictions are invariant to rotations, translations, and reflections.

    ICML 2024arXivCode
    BibTeX
    @inproceedings{nguyen2024structure,
  title         = {Structure-Aware {E(3)}-Invariant Molecular Conformer Aggregation Networks},
  author        = {Nguyen, Duy Minh Ho and Lukashina, Nina and Nguyen, Tai and Le, An Thai and Nguyen, TrungTin and Ho, Nhat and Peters, Jan and Sonntag, Daniel and Zaverkin, Viktor and Niepert, Mathias},
  booktitle     = {International Conference on Machine Learning (ICML)},
  series        = {Proceedings of Machine Learning Research},
  volume        = {235},
  pages         = {37736--37760},
  publisher     = {PMLR},
  year          = {2024},
  url           = {https://proceedings.mlr.press/v235/nguyen24g.html},
  eprint        = {2402.01975},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG}
}

  2. Dude: Dual Distribution-Aware Context Prompt Learning For Large Vision-Language Model

    D.M.H. Nguyen*, An Thai Le*, T.Q. Nguyen, N.T. Diep, T. Nguyen, D. Duong-Tran, Jan Peters, L. Shen, M. Niepert, D. Sonntag

    Adapts vision-language models from few examples by aligning shared and LLM-generated class prompts with visual tokens through unbalanced optimal transport.

    ACML 2024PMLRarXiv
    BibTeX
    @inproceedings{nguyen2024dude,
  title         = {{Dude}: Dual Distribution-Aware Context Prompt Learning For Large Vision-Language Model},
  author        = {Nguyen, Duy Minh Ho and Le, An Thai and Nguyen, Trung Quoc and Diep, Nghiem Tuong and Nguyen, Tai and Duong-Tran, Duy and Peters, Jan and Shen, Li and Niepert, Mathias and Sonntag, Daniel},
  booktitle     = {Asian Conference on Machine Learning (ACML)},
  series        = {Proceedings of Machine Learning Research},
  volume        = {260},
  pages         = {687--702},
  publisher     = {PMLR},
  year          = {2024},
  url           = {https://proceedings.mlr.press/v260/nguyen25c.html},
  eprint        = {2407.04489},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

  3. Grasp Diffusion Network: Learning Grasp Generators from Partial Point Clouds with Diffusion Models in SO(3)×R³ GDN

    João Carvalho, An Thai Le, Philipp Jahr, Qiao Sun, Julen Urain, Dorothea Koert, Jan Peters

    Generates grasp poses from a partial point cloud using diffusion on SO(3)×R³, with collision-cost guidance. Evaluated on grasping from raw depth observations.

    Preprint · 2024arXiv
    BibTeX
    @misc{carvalho2024grasp,
  title         = {Grasp Diffusion Network: Learning Grasp Generators from Partial Point Clouds with Diffusion Models in {SO(3)}$\times${R}$^3$},
  author        = {Carvalho, Jo{\~a}o and Le, An Thai and Jahr, Philipp and Sun, Qiao and Urain, Julen and Koert, Dorothea and Peters, Jan},
  year          = {2024},
  eprint        = {2412.08398},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

2023

  1. Accelerating Motion Planning via Optimal Transport MPOT

    An Thai Le, Georgia Chalvatzaki, Armin Biess, Jan Peters

    Optimizes batches of trajectories without gradients. The Sinkhorn Step coordinates waypoint updates using entropic optimal transport, with a Gaussian-process smoothness prior.

    NeurIPS 2023arXivCode
    BibTeX
    @inproceedings{le2023accelerating,
  title         = {Accelerating Motion Planning via Optimal Transport},
  author        = {Le, An Thai and Chalvatzaki, Georgia and Biess, Armin and Peters, Jan},
  booktitle     = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume        = {36},
  pages         = {78453--78482},
  publisher     = {Curran Associates, Inc.},
  year          = {2023},
  doi           = {10.52202/075280-3430},
  eprint        = {2309.15970},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  2. Hierarchical Policy Blending As Optimal Transport HiPBOT

    An Thai Le, Kay Hansel, Jan Peters, Georgia Chalvatzaki

    Uses unbalanced optimal transport in a look-ahead layer to choose weights for blending reactive Riemannian motion policies.

    L4DC 2023Proceedings
    BibTeX
    @inproceedings{le2023hierarchical,
  title         = {Hierarchical Policy Blending As Optimal Transport},
  author        = {Le, An Thai and Hansel, Kay and Peters, Jan and Chalvatzaki, Georgia},
  booktitle     = {Learning for Dynamics and Control Conference (L4DC)},
  series        = {Proceedings of Machine Learning Research},
  volume        = {211},
  pages         = {797--812},
  publisher     = {PMLR},
  year          = {2023},
  url           = {https://proceedings.mlr.press/v211/le23a.html},
  eprint        = {2212.01938},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  3. Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models MPD

    João Carvalho, An Thai Le, Mark Baierl, Dorothea Koert, Jan Peters

    Learns diffusion priors over robot trajectories and generates goal-conditioned motion plans through denoising. The earlier conference version of Motion Planning Diffusion.

    IROS 2023arXiv
    BibTeX
    @inproceedings{carvalho2023motion,
  title         = {Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models},
  author        = {Carvalho, Jo{\~a}o and Le, An Thai and Baierl, Mark and Koert, Dorothea and Peters, Jan},
  booktitle     = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages         = {1916--1923},
  year          = {2023},
  doi           = {10.1109/IROS55552.2023.10342382},
  eprint        = {2308.01557},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  4. Learning to reason over scene graphs: a case study of finetuning GPT-2 into a robot language model for grounded task planning

    Georgia Chalvatzaki, Ali Younes, Daljeet Nandha, An Thai Le, L.F.R. Ribeiro, I. Gurevych

    Fine-tunes GPT-2 on scene graphs to turn long-horizon requests into subgoals for a planner. Evaluated on ALFRED.

    Frontiers in Robotics and AI 2023arXiv
    BibTeX
    @article{chalvatzaki2023learning,
  title         = {Learning to Reason over Scene Graphs: A Case Study of Finetuning {GPT-2} into a Robot Language Model for Grounded Task Planning},
  author        = {Chalvatzaki, Georgia and Younes, Ali and Nandha, Daljeet and Le, An Thai and Ribeiro, Leonardo F. R. and Gurevych, Iryna},
  journal       = {Frontiers in Robotics and AI},
  volume        = {10},
  pages         = {1221739},
  year          = {2023},
  doi           = {10.3389/frobt.2023.1221739},
  eprint        = {2305.07716},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

2022

  1. Learning Implicit Priors for Motion Optimization StochGPMP

    Julen Urain*, An Thai Le*, Alexander Lambert*, Georgia Chalvatzaki, Byron Boots, Jan Peters

    Learns energy-based motion priors to initialize trajectory optimization or guide it through cost factors. Tested in simulation and on a real robot.

    IROS 2022arXivProject pageCode
    BibTeX
    @inproceedings{urain2022learning,
  title         = {Learning Implicit Priors for Motion Optimization},
  author        = {Urain, Julen and Le, An Thai and Lambert, Alexander and Chalvatzaki, Georgia and Boots, Byron and Peters, Jan},
  booktitle     = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages         = {7672--7679},
  year          = {2022},
  doi           = {10.1109/IROS47612.2022.9981264},
  eprint        = {2204.05369},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

2021

  1. Learning forceful manipulation skills from multi-modal human demonstrations

    An Thai Le, Meng Guo, Niels van Duijkeren, L. Rozo, R. Krug, A.G. Kupcsik, M. Bürger

    Learns insertion, sliding, and twisting from demonstrations of pose and force, using attractor-based impedance control. Demonstrated in e-bike motor assembly.

    IROS 2021arXiv
    BibTeX
    @inproceedings{le2021learning,
  title         = {Learning Forceful Manipulation Skills from Multi-modal Human Demonstrations},
  author        = {Le, An Thai and Guo, Meng and van Duijkeren, Niels and Rozo, Leonel and Krug, Robert and Kupcsik, Andras G. and B{\"u}rger, Mathias},
  booktitle     = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages         = {7770--7777},
  year          = {2021},
  doi           = {10.1109/IROS51168.2021.9636828},
  eprint        = {2109.04222},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

  2. Hierarchical Human-Motion Prediction and Logic-Geometric Programming for Minimal Interference Human-Robot Tasks

    An Thai Le, P. Kratzer, S. Hagenmayer, M. Toussaint, Jim Mainprice

    Combines hierarchical human-motion prediction with logic-geometric replanning to reduce interference in shared workspaces. Evaluated on MoGaze.

    IEEE RO-MAN 2021arXiv
    BibTeX
    @inproceedings{le2021hierarchical,
  title         = {Hierarchical Human-Motion Prediction and Logic-Geometric Programming for Minimal Interference Human-Robot Tasks},
  author        = {Le, An Thai and Kratzer, Philipp and Hagenmayer, Simon and Toussaint, Marc and Mainprice, Jim},
  booktitle     = {IEEE International Conference on Robot and Human Interactive Communication (RO-MAN)},
  pages         = {7--14},
  year          = {2021},
  doi           = {10.1109/RO-MAN50785.2021.9515539},
  eprint        = {2104.08137},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}

Humanoid demos

At VinRobotics, our researchers and engineers test learned locomotion, whole-body control, and physical interaction on real humanoids.

Compliant whole-body control

CompliantWBC uses proprioceptive force estimates to coordinate the arms, torso, and legs of the ~70 kg VR-M3 during contact. The demos show guided interaction and cooperative tasks; the paper's passivity analysis is local and assumes fixed stiffness.

Project page

Perceptive stair locomotion

VR-M3 (~70 kg) climbs real staircases of varying geometry at up to 0.6 m/s, including with a 5 kg payload. It uses a single onboard camera and a learned locomotion policy, with no LiDAR, motion capture, or teleoperation.

Read article

Fast walking on VR-H3

VR-H3 (178 cm, 85 kg) uses an RL locomotion policy trained with phase-aware gait rewards and domain randomization. The training curriculum increases difficulty only after the policy learns to balance at each level.

Read article

Built from the motor up

In-house actuators and real-time EtherCAT communication support learned locomotion on our humanoid platform. The team reports walking at 1.5 m/s, reaching 1.8 m/s with controllers trained on human motion data.

Read article

Global debut

Platform preview for Computex and ICRA 2026: whole-body teleoperation, dynamic payload handling, MPC + RL locomotion, and perception-action learning.

Read article

Join us

Our group is based at VinUniversity in Hanoi, with close ties to VinRobotics and IAS at TU Darmstadt. If you enjoy designing algorithms, understanding why they work, and testing them on hardware, I'd like to hear from you.

How we work together

I help students develop their own research judgment. We work through the mathematics and code together: which assumptions matter, and what evidence would change our minds? We turn rough ideas into questions we can test, with room to make mistakes and rethink. I encourage students to pursue their own directions and question mine.

Open positions and projects

How to apply

  • Email an@robot-learning.de with the subject [Join][<track>] <Your Name>.
  • Attach a CV and a transcript, and link to code you wrote.
  • Add a paragraph (under 300 words) about one of our papers: an idea that interests you, a question about the method, or an experiment you would like to try.

I read emails with this subject line first, but may not be able to reply to every message. Thanks for your patience.

Group

The students, residents, alumni, and collaborators behind our research. Read about how we work together and how to join.

MSc students

Dinh Van The Long Learning under partial observability

VinRobotics residents

Trinh Thi Cuc Whole-body compliance
Dang Truong Duy Payload-robust locomotion
Le Anh Chien Humanoid locomotion & control
Nguyen Viet Duong Neuromorphic learning
Ly Phuc Thanh Perceptive humanoid locomotion
Do Tan Dung Humanoid behavior models
Ho Thinh Hung VLA quantization & deployment
Ha Thien Loc Equivariant robot policies
Nguyen Quang Tan Equivariance & action generation
Ho Trong Bao Asynchronous robot policies
Nguyen Dang Khanh Efficient VLA inference
Le Van Minh World models & synthetic data evaluation
Nguyen Minh Hoang Reinforcement learning & agents
Dao Duc Thinh Egocentric VLA learning
Truong Minh Hoang CPU inference & SIMD kernels
Nguyen Truong Chinh VLA inference runtimes

Alumni

Phuong Tuan Dat Cross-embodiment robot learning
Magnus Dierking Contact-rich model predictive control
Caio Freitas Graph diffusion for imitation learning
Qiao Sun Geometric diffusion for grasping
Denis Andrić Learning Hamiltonian dynamics
Sebastian Zach Reactive motion generation

Collaborators

Zachary Kingston Purdue University
Meng Guo Peking University
Nguyen Ho Minh Duy DFKI
Ngo Anh Vien VinRobotics
Jan Peters TU Darmstadt
Georgia Chalvatzaki TU Darmstadt
João Carvalho DFKI
Viet T. Nguyen University of Würzburg

Open source

Research code and deployment tools developed with students and collaborators, including our releases at CAIR, VinUniversity. Start with GTMP or MPOT for planning, vla.simd for CPU policy inference, or warpax for numerical relativity. Questions and bug reports are welcome; reproducible examples help us investigate.

cair-vinuni/vla.simd C++ inference for VLA policies with native SIMD kernels on x86, Apple Silicon, and Raspberry Pi 5. C++240 cair-vinuni/FoldQuantVLA Native INT4 and INT8 inference for VLA models through consistent offline folding, with deployment on Jetson AGX Orin. Python242 cair-vinuni/FOCA Adapts VLA policies from few demonstrations by learning from future interactions and goals. ICML 2026. Python922 cair-vinuni/CLP_VLA Prunes redundant VLA layers before fine-tuning, with training and evaluation scripts for robot manipulation. Jupyter Notebook991 cair-vinuni/RoboGaze Diagnoses failures in videos generated by robot world models using structured vision-language analysis. Python1511 cair-vinuni/paint-action-chunking Project materials for PAINT: connecting action chunks through initial noise selection without retraining. CoRL 2026. 140 VinRobotics/vla.cpp C++ runtime for onboard inference with eleven vision-language-action models. C++19834 anindex/mpot Gradient-free trajectory optimization with optimal transport in PyTorch. NeurIPS 2023. Python7111 anindex/mtp Global and local trajectory sampling for model predictive control in JAX. TMLR 2025 & ICLR 2026. Python442 anindex/gtmp Parallel motion planning with layered graphs and tensor operations in JAX. RA-L 2025 & ICRA 2026. Python405 anindex/polystep Optimal-transport training for non-differentiable networks using only forward evaluations. TMLR 2026. Python331 anindex/warpax Vectorized auto-diff in JAX for checking energy conditions in warp drive spacetimes. CQG 2026. Python31 anindex/penrose_process Rocket-flyby simulations testing energy extraction and escape around rotating black holes. Phys. Rev. D 2026. Python20 anindex/aac Learned compression of landmark heuristics for A* search, preserving admissibility by construction. Python40 VinRobotics/vinrobotics_mjlab RL training pipeline for high-payload humanoid locomotion, built on MuJoCo Warp. Python449 VinRobotics/model-quantization-recipes Quantization recipes for language and speech models, with deployment checks. Python356 anindex/note_model_opt Notes and runnable examples on model deployment. 272 vincekurtz/hydrax (contributor) Sampling-based model predictive control on the GPU with JAX/MJX. Includes MTP. Python30453

Teaching & mentoring

I start with a small example, work through the mathematics, then implement it to test the assumptions. Our student projects and mentoring follow the same approach: learn to ask precise questions and judge the evidence for yourself.

Courses at TU Darmstadt

Open learning resources

Community & service

Editorial and reviewing roles

Associate Editor
IEEE RA-L, ICRA 2027
Area Chair
ICLR 2027, CoRL 2026, RLC 2024
Conference reviewer
IROS, ICRA, RSS, L4DC, NeurIPS (Top Reviewer 2026), ICML, ICLR, AAAI
Journal reviewer
IEEE RA-L, IEEE T-RO, Neurocomputing, TMLR, Frontiers in Robotics and AI

Selected invited talks

Experience & education

  1. Oct 2025–present
    Hanoi, Vietnam

    Assistant Professor

    VinUniversity

    Building a research group on efficient learning and planning for robot locomotion and manipulation.

  2. Oct 2025–present
    Darmstadt, Germany

    Visiting Professor

    TU Darmstadt

    Co-advising MSc and PhD students in robot learning at IAS.

  3. Aug 2025–present
    Hanoi, Vietnam

    Director of Foundation AI

    VinRobotics
    • RL stack for high-payload humanoid locomotion
    • Humanoid VLA architecture and training recipe
    • Model optimization and edge-deployment toolchain
  4. 2022–2025
    Darmstadt, Germany

    PhD in Computer Science

    TU Darmstadt, Intelligent Autonomous Systems (IAS)

    Thesis: Tensor Search Methods for Vectorizing Motion Planning, supervised by Prof. Jan Peters.

  5. 2019–2021
    Stuttgart, Germany

    MSc in Information Technology

    Universität Stuttgart

    Thesis: Learning task-parameterized Riemannian motion policies, supervised by Dr. Jim Mainprice and Dr. Meng Guo. Graduated first class. Info-Preis for the best diploma, Sony Research Award, Deutschlandstipendium.

  6. May–Dec 2020
    Renningen, Germany

    Research Intern

    Bosch Center for Artificial Intelligence

    Forceful imitation learning for e-bike assembly, hosted by Dr. Meng Guo in the robotics team.

  7. Nov 2019–Apr 2020
    Stuttgart, Germany

    Research Assistant

    HLRS (High-Performance Computing Center Stuttgart)

    Configured and maintained HPC systems.

  8. 2015–2019
    Frankfurt, Germany

    BEng in Electrical Engineering and Information Technology

    Frankfurt University of Applied Sciences

    Thesis: Approaches to solve the kidnapped robot problem. Graduated first class. DAAD, AmCham and eSilicon scholarships.

  9. Jan–May 2017
    Ho Chi Minh City, Vietnam

    Engineering Intern

    Intel Corporation

    Designed systems to analyze high-volume manufacturing unit-test data and supported quality validation and reporting on the Thunderbolt line.