ACT policy for the SO-ARM101, trained by Φ (Physical Hardware Intelligence), the robotics group
at Northeastern University's Silicon Valley campus.
Part of an object-holdout generalization study: train on two of three objects, hold the third
out entirely, and measure whether grasp behaviour transfers to an object size never seen in training.
Held out of training: red cube, episodes 0-39 (25 mm).
eval_loss is pure L1 with the CVAE latent set to zero, computed on the 40 held-out episodes
every 10,000 steps. It is not the training objective: in eval mode ACT skips the VAE encoder, so no
KL term is included, which makes these numbers comparable across kl_weight settings.
⚠️ The final checkpoint is not the best one. Training loss reached ~0.053 on every run in this
study while held-out loss varied by up to 28%. Pick a checkpoint by held-out loss, not by the last
step.
Only that checkpoint is downloaded, not the whole repo. Any of
020000 040000 060000 080000 100000 is valid.
Caveats
No hardware success rate yet. These numbers are held-out L1, not scored rollouts.
The two holdout splits are not comparable to each other: different held-out sets (27,272 vs
18,277 frames), different objects, different episode lengths.
Camera keys in the source dataset are correct and need no transposition.
