Projects
Things I've built, contributed to, broken, and occasionally got working on real robots.
Featured Work
PAL Robotics · July 2025 – Present
Built pal_mjlab (PAL's mjlab extension) and ROS deployment package, enabling org-wide RL development.
Results
- Velocity Tracking: RL-based walking controller to replace legacy MPC stack
- Motion Imitation: GMR retargeting + BeyondMimic for natural movement imitation
- Fall Recovery: Trained recovery policies (sim-validated, hardware deployment pending)
mjlab
PyTorch
ROS
MuJoCo
PPO
Open Source · Led by Kevin Zakka
Core contributor to mjlab, a MuJoCo Warp-based RL framework built by Kevin Zakka.
Results
- ~2× faster than Isaac Lab in object-rich manipulation environments
- Successful sim-to-real transfer with near zero domain randomization
- Adopted by UC Berkeley's graduate robotics course (ME 292b/193b)
MuJoCo Warp
Python
PyTorch
Open Source · Personal Project
End-to-end pipeline for training RL manipulation policies in Isaac Lab and deploying them on a real Kinova Gen3 arm via ROS 2. Created as the missing "Hello World → Real Robot" bridge for Isaac Lab users.
Results
- 156k+ social impressions (LinkedIn, X, Reddit)
- Adopted by INIT Lab (ÉTS Montréal) and researchers in Europe and Japan
- Includes docs, tutorials, and pre-trained weights
Isaac Lab
ROS 2
PyTorch
Python
University of Tokyo · June – Sept 2024
Built a ROS pipeline fusing Velodyne LiDAR with RealSense D455 for semantic navigation on a quadruped platform. Enabled object-centric SLAM and person-following for Jiaxu Wu's PhD research.
Results
- Live colored point-cloud at 10 Hz with YOLO object labels
- Became the lab's reference template for LiDAR-camera fusion
- Resolved ARM-Docker issues on Jetson Xavier for edge deployment
ROS
C++
YOLO
CLIP
Docker
Other Projects
Open Source · Collaboration
RL-based manipulation tasks for SO-ARM100/101 robots using Isaac Lab. Presented at ROSCon España (Nov 2025).
Isaac Lab
PyTorch
ROS 2
Personal Project · ROS 2
Designed, 3D-printed, and brought up a 6-DOF wheeled biped robot. Wrote custom controllers and RViz visualization to validate stability and zero-cantilever design.
ROS 2
SolidWorks
Python
Naova RoboCup Team · Nov 2024 – April 2025
Enhanced NAO biped walking stability by ~15% by converting discrete to continuous control signals. Also developed a neural network for automatic referee signal interpretation.
Python
PyTorch
Control