Current focus / October 2026
Rough-terrain walking, evaluated in simulation.
I finished training and evaluating a rough-terrain walking policy for the Pioneer humanoid. Alongside it, I built gait diagnostics, refined contact sensing, and prepared a complete stairs environment for the next stage of training.
The role
My work starts before the first training run.
I build the simulation and reinforcement-learning workflow behind learned locomotion for the team's custom humanoid.
Before training, I validate the robot model, joint behavior, actuator dynamics, and contact sensing. I then build the Isaac Lab environments and shape rewards around the failures each policy exposes.
I train with PyTorch and RSL-RL across GPU-parallel environments, build focused diagnostics to separate model, physics, and policy issues, and work with the humanoid team to keep simulation aligned with the team's mechanical model.
Projects
Selected work.
- 01 September to October 2026 Pioneer Humanoid Locomotion Trained and evaluated rough-terrain walking, refined contact sensing, and built gait diagnostics and a reusable stairs training environment.
- 02 June to July 2026 GPU-Accelerated Humanoid Locomotion Converted the custom 12-DOF lower body into a trainable Isaac Lab system and produced a repeatable walking policy from a fresh PPO run.
- 03 May to September 2026 Lidar-to-Control Navigation Built four C++ ROS 2 nodes for lidar mapping, A* planning, and closed-loop path following in Gazebo.
Working with
Python / PyTorch / NVIDIA Isaac Lab / RSL-RL / PPO / CUDA / C++ / ROS 2 / Gazebo / Foxglove