I want humanoids that can take a task, however it's specified, and carry it out in whatever scene they're in. My research is on whole-body control grounded in perception, where the task's intent stays fixed and the execution adapts to what the robot sees.
At VinRobotics I lead the Reinforcement Learning Department. We've taken perceptive locomotion and whole-body compliance from simulation onto real humanoids under load, and I'm now building perception into the pretraining of general whole-body controllers.
I did my MSE in Robotics at Penn and a structured-light vision thesis at HCMUT. I'm applying to PhD programs in robot learning, and I'm open to full-time roles and research collaborations in the meantime.
Outside of work, I travel a lot. More in Travel.
Last touched October 2026
SeeBFM trains BFM-Zero further with a height scan of the ground around the robot. A reward, a motion or a typed sentence becomes one prompt, and the policy reads the ground to carry it out. The network that turns a reward into a prompt stays blind to the terrain, so one reward gives the same prompt on every ground. On the robot, a student network predicts the scan from two depth cameras.
Project page → Playground →
SOFIE learns perceptive humanoid locomotion entirely offline, from rollouts of experts of mixed quality. It scores whole footsteps by their advantage and imitates the better ones more, so one policy walks three humanoids over stepping stones, stairs and rough ground where behavior cloning copies the mistakes. Fine-tuned on a small dataset from an unseen humanoid, a policy pretrained with SOFIE also beats training from scratch.
Project page →
Model-based RL agents usually tune the planner and the critic's learning target separately. Across ten tasks, TD-MTP finds that the two interact. On a humanoid stair task, each change alone lowers the return, and the two combined beat the baseline.
Project page →
Perceptive humanoid locomotion without the usual motion-capture and distillation pipeline. A terrain classifier routes a five-expert mixture over a sensed elevation grid, and every foothold target is refined by closed-form gradient steps on an interpolated terrain-cost field rather than snapped to a fixed grid. Crosses an 85 m five-segment waypoint course 18/20 end-to-end in a second, deployment-fidelity simulator.
Project page →
A perceptive locomotion system for payload-robust humanoid walking. Multi-channel terrain cost maps guide a GPU-parallel foothold planner, with virtual wrench injection enabling payload-compliance training without force sensors. Achieves 1.0 m/s on stairs with 20 cm risers and zero-shot transfer to a physical humanoid carrying up to 20 kg.
Project page →
Whole-body compliance for a 70 kg heavy humanoid under high-payload and external-force scenarios. A multi-site impedance reference controller, variational force latent encoder, and bounded residual mechanism jointly edit impedance equilibria, validated on surface wiping, force reaction, loaded squatting, and cooperative transport with a 0.98 success rate.
Project page →