I'm an Assistant Professor at VinUniversity, a Visiting Professor at IAS, TU Darmstadt, and Director of Foundation AI at VinRobotics, where my team builds the RL stack for high-payload humanoid locomotion, humanoid VLA architectures and training recipes, and the model optimization and edge-deployment toolchain.

I study how robots can plan, learn, and act reliably with limited computation and data. My group works on three directions that increasingly fit together:

• Parallel planners. We represent candidate plans and their costs as tensors so GPUs can search many of them at once (tensor search), and pair this structure with optimal transport and diffusion / flow-matching priors for long-horizon, multimodal, non-convex problems.
• Scaled-down robot foundation models. VLA policies that adapt from fewer demonstrations and run on affordable onboard computers: native low-bit quantization, a C++ inference runtime, and shared SIMD kernels that run policies on CPUs down to a Raspberry Pi 5.
• Whole-body control through contact. Force-aware controllers that let 70 to 85 kg humanoids respond to contact and changing payloads with their arms, torso, and legs together.

Next up: whole-body VLAs trained on contact-rich data, with planners checking their actions and compliant controllers executing them.

On weekends, I write JAX/PyTorch simulators for curved spacetimes (Kerr black holes, the Penrose process, warp-drive energy conditions), because these physics are art!

• Model Tensor Planning, TMLR 2025 / ICLR 2026 (J2C)
• Global Tensor Motion Planning, IEEE RA-L 2025 / ICRA 2026
• Anytime Global Tensor Motion Planning, under review
• Motion Planning Diffusion, IEEE T-RO 2025 / AAAI 2026 (journal track) / IROS 2023
• Accelerating Motion Planning via Optimal Transport, NeurIPS 2023
• CLOT: Multi-Robot Motion Planning via Collaborative Optimal Transport under STL Tasks, ICRA 2026

• FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding, under review
• vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation, under review
• vla.cpp: A Unified Inference Runtime for Vision-Language-Action Models, under review
• FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation, ICML 2026
• Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection, CoRL 2026
• Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think, under review

• CompliantWBC: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets, under review
• TACT-ful: Multi-Channel Terrain Affordance and Compliance Training for Payload-Robust Perceptive Humanoid Locomotion, under review
• DoublyAware: Dual Planning and Policy Awareness for Temporal Difference Learning in Humanoid Locomotion, IEEE RA-L 2026

• Training Non-Differentiable Networks via Optimal Transport, TMLR 2026
• AAC: Admissible-by-Architecture Differentiable Landmark Compression for ALT, under review

• Rarity of rocket-driven Penrose extraction in Kerr spacetime, Physical Review D 2026
• Observer-robust energy condition verification for warp drive spacetimes, Classical and Quantum Gravity 2026

Full list on my website and Google Scholar.

• 🌐 Website: anindex.github.io
• 🎓 Google Scholar: An Thai Le
• 🐦 X: @an_thai_le
• 💌 Email: an@robot-learning.de