tavishh/robotarm

ACT (CVAE) — SO-101 cubes & cylinder pick-and-place

Team TRON submission for the Φ (Physical Hardware Intelligence) ACT training competition — Northeastern SV robotics SIG.

Summary

ACT policy (with CVAE) trained on the phi_so101_cubes_cylinder_v1 dataset (120 episodes, 66,873 frames, 3 cameras) to pick up a cube or cylinder and place it in a box. Checkpoint submitted is step 60,000 of 100,000, chosen by held-out loss rather than the final step.

• Held-out L1 (step 60,000): 0.500 — the lowest of any checkpoint across two training runs (this CVAE run and a behavioral-cloning baseline run without the CVAE latent), evaluated on the same 30-episode position holdout.
• The final checkpoint (step 100,000) scored 4.8% worse (0.524) — the run had already begun overfitting past step 60,000–80,000.
• A parallel behavioral-cloning baseline (use_vae=false, no latent) peaked at step 80,000 with L1 = 0.513 — still worse than this checkpoint.