Montreal Robotics Summer School trains Unitree Go1 to autonomously hunt its food bowl
GlenBerseth · x · 2026-09-09
At the 2026 Montreal Robotics Summer School (Aug 9–14, Mila), graduate student teams trained and deployed policies on a real Unitree Go1 quadruped in a challenge called "Bonne Appétit — Feed the Robot Dog."
- The week-long program covered state estimation, perception, manipulation, planning, and sim2real reinforcement learning, culminating in a live robot challenge.
- Three scoring tiers: joystick-controlled straight-line walk (1 pt), joystick-controlled obstacle course (2 pts), and fully autonomous navigation to the bowl using only a front-facing monocular fisheye camera and AprilTag detection (3 pts).
- The core technical pipeline mirrored real robotics practice: train a robust low-level walking policy in Isaac Lab simulation, transfer it to the physical Go1, then layer vision-based navigation on top.
- Overall winners: Team 5 (Mahsa Hasheminejad, Michael Jenson, SayedHamid Bahreini, Zekai Jin) and Team 6 (Mohamed Nabail, Mohamed Samir, Rodrigo Murillo Aranda).
More from Embodied
- Lightwheel open-sources 100,000 hours of egocentric video for robot training — chris_j_paxton · 2026-09-09
- NVIDIA Cosmos powers 4 of top 5 winners in ECCV AI City Challenge forecasting track — NVIDIAAI · 2026-09-09
- Chinese suppliers shipped parts for 5,000 Optimus robots, 15,000 targeted by year end — JOBhakdi · 2026-09-09
- Dollhouse Previews Miniverse: Control Embodied AI Models in Your Browser, SOMA47 Climb Tracker Demo — carlosdponx · 2026-09-09
- Dollhouse Launches Miniverse: Share Robot Simulations That Run in the Browser, No Weights Exposed — carlosdponx · 2026-09-09
- Arm gathers 80+ companies for Physical AI initiative with Robotics Capability Framework — pdamodaran · 2026-09-09