omnireset-easybhome-bases

OmniReset easyBhome base diffusion policies

State diffusion policies (BC) for the three OmniReset easyBhome tasks, trained 2026-08-18 on the 200-episode expert datasets at knightnemo/omnireset-easybhome-demos (1000 epochs, batch 256, lr 1e-4, episode-level val split). Each {task}/base.pt is the training snapshot that ranked best by evaluated environment success (never by validation loss — the val-best checkpoint scores 2.5% on peg).

Success rates

One evaluation protocol: the RL (RlAligned) environment — OmniReset's own reward and terminations — at the 320-step episode protocol, over one wave of 1024 fresh-boot environments (±1.5% binomial at these rates).

Fleet-scale measurement caveat: UWLab's ProgressContext.reset (as of commit 36d98af) zeroes every environment's hold counter on any partial reset. At 1024+ environments some environment resets nearly every step, so the hold-for-5-steps success criterion can never fire: the drawer base reads 0.0% while ~57% of environments sit stably inserted. Evaluating or training these policies at fleet scale requires patching that reset to be per-environment (continuous_success_counter[env_ids] = 0); the numbers above use the patched criterion. At ≤128 environments the bug is nearly invisible.

Per-task protocol evaluation records ({task}/base_eval_protocol.json) and training metrics are included alongside each policy.

Payload format: base_policy.dp.state_diffusion_policy.v1, loaded by base_policy.furniture_loader.load_furniture_state_policy in knightnemo/Understand_Fleet_RL.