# An Thai Le

> Robotics and machine learning researcher who turns sequential robot algorithms into parallel ones and fits robot foundation models onto small, low-cost hardware. Assistant Professor at VinUniversity (Hanoi, Vietnam), Visiting Professor at TU Darmstadt (IAS, Jan Peters' lab), and Director of Foundation AI at VinRobotics, where the methods run on humanoid robots. Contact: an@robot-learning.de.

- **Research to deployment.** Combines academic research with industry leadership. At VinRobotics: leads the RL stack for high-payload humanoid locomotion, the humanoid VLA architecture and training recipe, and the edge-deployment toolchain.
- **Parallel motion planning.** Recasts planning as tensor search that runs on GPUs: Global Tensor Motion Planning (IEEE RA-L 2025, ICRA 2026), Model Tensor Planning (TMLR 2025, ICLR 2026), Accelerating Motion Planning via Optimal Transport (NeurIPS 2023), Motion Planning Diffusion (IEEE T-RO 2025).
- **Efficient robot foundation models.** VLA policies that run on CPUs, Raspberry Pi 5 and Jetson Orin (vla.simd, FoldQuantVLA); data-efficient VLA adaptation (FOCA, ICML 2026); smooth asynchronous execution of action chunks without retraining (PAINT, CoRL 2026).
- **Humanoids in contact.** Whole-body compliance for a 70 kg humanoid (CompliantWBC), payload-robust perceptive stair descent (TACT-ful), uncertainty-aware humanoid locomotion (DoublyAware, IEEE RA-L 2026).
- **Trusted by peers.** Associate Editor for IEEE RA-L and ICRA 2027; Area Chair for ICML 2027, ICLR 2027, CoRL 2026 and RLC 2024; NeurIPS 2026 Top Reviewer.
- **Range.** Gradient-free training through optimal transport (PolyStep, TMLR 2026), and differentiable JAX tools for general relativity published in Physical Review D and Classical and Quantum Gravity.
- **Mentor.** Advises MSc students and VinRobotics residents; student-led work includes PAINT (CoRL 2026).
- **Background.** PhD (2025) at TU Darmstadt with Jan Peters, thesis *Tensor Search Methods for Vectorizing Motion Planning*, external examiner Siddhartha Srinivasa. MSc Universität Stuttgart; research at Bosch Center for Artificial Intelligence and HLRS; engineering at Intel.
- **Good fit for:** collaborations, invited talks and reviews on GPU-parallel planning, efficient VLA inference and quantization, humanoid locomotion and compliance, and optimal transport in robotics and ML; MSc and PhD supervision; industry partnerships that deploy robot learning on real hardware. Students and collaborators: https://anindex.github.io/#join
- Name variants: An Thai Le, An T. Le, Le Thai An. ORCID: 0000-0003-0929-3316.
- Paper status (accepted, under review, ongoing) is kept current on the homepage. Cite with the BibTeX entries in llms-full.txt.

## Profile

- [Homepage as Markdown](https://anindex.github.io/index.md): bio, news, research threads, every publication with status and links, group members, open source, teaching, service, experience
- [Full text with BibTeX](https://anindex.github.io/llms-full.txt): the Markdown homepage plus a BibTeX entry for every paper
- [News feed (Atom)](https://anindex.github.io/feed.xml): new papers, roles and talks
- [Homepage (HTML)](https://anindex.github.io/)

## Profiles elsewhere

- [Google Scholar](https://scholar.google.com/citations?user=k0r0tfUAAAAJ)
- [ORCID](https://orcid.org/0000-0003-0929-3316)
- [OpenReview](https://openreview.net/profile?id=~An_Thai_Le1)
- [GitHub](https://github.com/anindex): personal research code
- [CAIR, VinUniversity on GitHub](https://github.com/cair-vinuni): group releases
- [Hugging Face](https://huggingface.co/anindex)
- [LinkedIn](https://www.linkedin.com/in/anindex/)
- [X](https://x.com/an_thai_le)
- [VinUniversity profile](https://vinuni.edu.vn/people/le-thai-an/)
- [IAS, TU Darmstadt profile](https://www.ias.informatik.tu-darmstadt.de/Team/AnThaiLe)

## Selected publications

- [CompliantWBC: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets](https://compliantwbc.github.io/): Under review · 2026. 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.
- [FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding](https://arxiv.org/abs/2609.24433): Preprint · 2026. 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.
- [vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation](https://arxiv.org/abs/2609.24274): Preprint · 2026. 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.
- [Training Non-Differentiable Networks via Optimal Transport](https://openreview.net/forum?id=8mlcqTTMuU): TMLR 2026. 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.
- [Observer-robust energy condition verification for warp drive spacetimes](https://doi.org/10.1088/1361-6382/aeb25a): Class. Quantum Grav. 2026. 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.
- [Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection](https://arxiv.org/abs/2606.19774): CoRL 2026. 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.
- [FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation](https://icml.cc/virtual/2026/poster/66754): ICML 2026. 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.
- [Model Tensor Planning](https://openreview.net/forum?id=fk1ZZdXCE3): TMLR 2025, ICLR 2026 (J2C). Combines global candidates from layered graphs with local trajectory sampling for model predictive control. Batches simulator rollouts in MuJoCo XLA; available in hydrax.
- [Global Tensor Motion Planning](https://sites.google.com/view/gtmp/home): IEEE RA-L 2025, ICRA 2026. Searches a layered graph with tensor operations to produce batches of smooth spline paths in parallel. Provides probabilistic completeness under the paper's assumptions.
- [Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models](https://sites.google.com/view/motionplanningdiffusion): IEEE T-RO 2025, AAAI 2026 (journal track). 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.
- [Accelerating Motion Planning via Optimal Transport](https://sites.google.com/view/sinkhorn-step/): NeurIPS 2023. Optimizes batches of trajectories without gradients. The Sinkhorn Step coordinates waypoint updates using entropic optimal transport, with a Gaussian-process smoothness prior.

## Latest news

- 2026-10: Happy to serve as an Associate Editor for IEEE Robotics and Automation Letters (RA-L) and an Area Chair for ICML 2027. I look forward to handling your submissions!
- 2026-09: Start Right, Arrive Right (PAINT) has been accepted at CoRL 2026. PAINT selects initial noise to smoothly connect action chunks during asynchronous execution, without retraining. Congratulations to my student Ho Trong Bao and the team!
- 2026-09: Happy to serve as an Associate Editor for ICRA 2027 and an Area Chair for ICLR 2027. I look forward to reading and discussing your work.
- 2026-09: Observer-robust energy condition verification for warp drive spacetimes (warpax) has been accepted by Classical and Quantum Gravity (published online October 2026). The paper introduces vectorized auto-diff tools with JAX for checking energy conditions in warp drive spacetimes. Code is online.
- 2026-09: Training Non-Differentiable Networks via Optimal Transport (PolyStep) has been accepted by TMLR. PolyStep trains non-differentiable networks through optimal transport with only forward evaluations. Code is online.

## Talks

- [Designing Planning Algorithms for the Era of Parallelism](https://anindex.github.io/slides/roboarch/): invited talk slides, ICRA 2026 Workshop on Robot Architecture; tensor search and parallel motion planning

## Optional

- [PhD thesis](https://tuprints.ulb.tu-darmstadt.de/entities/publication/41c837ff-8bab-4ad1-acaf-4d4602f41b26): Tensor Search Methods for Vectorizing Motion Planning (TU Darmstadt, 2025)
- [Model deployment notes](https://github.com/anindex/note_model_opt): runnable examples on model optimization and deployment