RoboCasa Wave1 — last tasks
Protocol: literal-5 seeds 10000..10004, -n 1 --n_test 50, OAT8, BoN N=8 vote.
Δ = mean(BoN)−mean(base); SEM_Δ = sqrt(SEM_b²+SEM_m²).
Table P (Wave1)
| task | baseline mean±SEM | BoN N=8 | Δ_BoN±SEM_Δ | status |
|---|---|---|---|---|
turn_off_sink_faucet |
52.4±2.5% | 56.0±3.4% | +3.6±4.2 pp | DONE |
turn_off_microwave |
(pending) | (pending) | — | BoN RUNNING — watcher will finalize |
Sink per-seed SR (%)
- baseline:
{'10000': 52.0, '10001': 58.0, '10002': 58.0, '10003': 46.0, '10004': 48.0} - BoN8:
{'10000': 54.0, '10001': 58.0, '10002': 64.0, '10003': 60.0, '10004': 44.0}
Artifacts
robocasa/<task>/
summary_literal5.json
baseline_seed{10000..10004}/eval_log.json
bon_n8_seed{10000..10004}/eval_log.json
Ckpts not in this repo (~0.9GB) — see hackhackhack66666/rc-model-two-last.