CHL0e/robomme-mfm

MFM agent on RoboMME: evaluation results

Results and logs of the MFM (Multimodal Filesystem Memory) agent on RoboMME (ICML 2026), a benchmark of memory in robot manipulation: 16 table-top Franka Panda tasks in ManiSkill3/SAPIEN, in four suites (counting, permanence, reference, imitation).

The agent is GPT (gpt-6.1-sol, via the Codex CLI) driving the robot through end-effector waypoints. Its memory is a working directory: an append-only log of every executed action, the frames each action started from and saw, the task's demonstration video, and notes the model keeps for itself. Code: mfm_robo/robomme_agent.py and scripts/robomme/ in the MFM-robo repository.

Runs

Run	Episodes	Successes	Overall
dev/ (2026-10-20)	16 tasks x test episodes 0, 1 (32)	28	87.5%
pilot/ (2026-10-20)	VideoUnmask, PickXtimes, episode 0 (2)	2	100%

"Overall" is the mean of the per-task success rates, as on the RoboMME leaderboard. With two episodes per task every task rate is 0, 50 or 100%, so these numbers show the setup works; they are not a benchmark result.

dev/ per task

Suite	Task	MFM (2 episodes)	GPT-6-Astra (waypoint, 50)	PonderPounce 9B (joint, 50 x 3)
Counting	BinFill	50	98	50.7
Counting	PickXtimes	100	100	93.3
Counting	SwingXtimes	100	98	72.0
Counting	StopCube	50	54	82.7
Permanence	VideoUnmask	100	94	94.7
Permanence	VideoUnmaskSwap	100	96	25.3
Permanence	ButtonUnmask	100	100	97.3
Permanence	ButtonUnmaskSwap	100	98	34.0
Reference	PickHighlight	100	98	85.3
Reference	VideoRepick	50	100	40.7
Reference	VideoPlaceButton	50	100	83.3
Reference	VideoPlaceOrder	100	98	79.3
Imitation	MoveCube	100	54	81.3
Imitation	InsertPeg	100	92	6.0
Imitation	PatternLock	100	100	12.7
Imitation	RouteStick	100	94	34.7
	Overall	87.5	92.1	60.8

Reference columns are the published leaderboard entries in the benchmark repository's doc/submission/.

The four failures:

• BinFill ep0: the red cube bounced out of the bin; the agent then put in the nearest cube without checking its color.
• StopCube ep1: it counted three arrivals and estimated the fourth at control step 210; its press reached the button at step 217, and the cube did not stop on the target.
• VideoRepick ep0: all cubes are green and two swap places late in the 257-frame video; the agent picked the cube at its original position.
• VideoPlaceButton ep0: it misread the 748-frame video and chose the wrong target.

The two video failures share a cause: the model sees each video as at most 32 evenly spaced frames, which can miss short events such as a swap.

Protocol

RoboMME's leaderboard settings: the test split, the waypoint action space (an absolute end-effector pose plus open/close per action, executed by the benchmark's motion planner), and max_steps=1300 control steps after the demonstration video. An environment error ends the episode as a failure. The agent observes front and wrist RGB, the end-effector pose and gripper state, the camera intrinsics and extrinsics, and front depth: no object state, oracle subgoals or multi-choice options. An episode also ends, as a failure, if the agent declares it done early, after 3 turns in a row without a usable plan, or after 150 model turns.

Benchmark code: RoboMME/robomme_benchmark at commit bf6fd29.

Differences from an official run:

• Rendering ran on CPU with Mesa lavapipe, not on a GPU. Frames differ from GPU frames only in shading (mean 5/255 per pixel); physics is identical.
• Five dev episodes (SwingXtimes ep0 and ep1, ButtonUnmask ep0, VideoPlaceOrder ep0, InsertPeg ep0) first crashed in the renderer before the agent acted. They were rerun after the fix, and the files here are from the reruns.

Layout

<run>/results.md, results.json     per-task and per-episode results, tokens, wall clock
<run>/jobs.txt, agent_args.txt     the episodes run and the agent flags
<run>/<Task>_ep<N>.log             the job's stdout: each executed action and its result
<run>/<Task>_ep<N>/summary.json    outcome, how it ended, control steps, tokens, timing
<run>/<Task>_ep<N>/logs.txt        the agent's memory log: action, result, state per entry, its plans
<run>/<Task>_ep<N>/notes/working.md  the agent's own notes
<run>/<Task>_ep<N>/rollout.mp4     demonstration video (red border) then the episode, front | wrist

Frames, motion frames, depth and Codex transcripts were not uploaded.