Previous videos about the simulation round: ▶️ Part 1 (RL loop, AWR + RECAP): ▶️ Part 2 (reward, advantage, inference-time optimization):

Full video (Part 2 of 3): 🎥 Coming next: Part 3: sim-to-real, DAgger and human-in-the-loop, the real-robot final at ICRA

The thread running through all of it is image augmentation, and it stays oddly underutilized in robotics. Randomize colors, lighting, camera pose, and physical properties, and the policy stops…

Full video (Part 1 of 3): RL theory to working pipeline, ~all of the above in detail: 🎥 Coming next: Part 2: reward engineering, advantage computation, inference-time optimization (CFG, best-of-N,…

Takeaway: even though I didn't properly ablate my solution, I think the overall approach is promising and I am sharing all the details here for the community to play with. The biggest improvement…