CHL0e/robocasa-eval-results

RoboCasa MFM agent evaluation results

Evaluation runs of the MFM agent (backend model gpt-6.1-sol) on the RoboCasa lifelong_phase4 task set: 18 kitchen tasks × 3 seeds, mobile manipulator (Franka Panda arm on a holonomic base with a torso lift).

run	variant	episodes finished	successes	episode success rate
lifelong_phase4/	default	54 / 54	2 (AlignSilverware s0, ClearSinkArea s1)	0.037
lifelong_phase4_code/	--variant code --privileged full --room none	13 / 54 (partial snapshot, 2026-10-19)	1 (ClearSinkArea s0)	0.077

Per-task tables, token and cost figures, and the model's own notes are in each run's results.md; machine-readable numbers are in results.json.

Layout

<run>/
  results.md, results.json     aggregate results
  jobs.txt, slurm/             job list and SLURM stdout/stderr
  <Task>_s<seed>.log           per-episode agent log
  <Task>_s<seed>/
    summary.json               task, seed, success, sim steps, horizon, wall time
    rollout.mp4                episode video
    index.jsonl, episodes/     per-entry memory (actions and outcomes)
    frames/                    camera views shown to the model
    codex_turns/               model turns
    logs.txt, last_message.txt, notes/, actions.json
    AGENTS.md, tools/          agent instructions and tools for that episode

Episodes in lifelong_phase4_code/ without a summary.json were still running when this snapshot was uploaded.