ishneet/so101-pairwise-eval

SO-101 pairwise policy eval

A blind pairwise evaluation, RoboArena-style, of three SO-100/101 checkpoints on one real SO-101 follower arm:

Model	Checkpoint
MolmoAct2	allenai/MolmoAct2-SO100_101
π0.5	hqfang/pi05-so100_101
GR00T N1.7	hqfang/groot-so100_101

The setup was 18 scenes (6 table setups × 3 lighting levels), with 10 tasks per scene and one model pair per scene. That gives 90 head-to-head comparisons and 180 rollouts, all scored with partial credit.

Results

Pair	Wins	Wins	Ties
MolmoAct2 vs GR00T	MolmoAct2 18	GR00T 1	11
MolmoAct2 vs π0.5	MolmoAct2 3	π0.5 1	26
π0.5 vs GR00T	π0.5 17	GR00T 1	12

Rank	Model	Bradley-Terry (ties = ½ win)	Mean partial score (n = 60)
1	MolmoAct2	1.61	21.2
2	π0.5	1.42	15.1
3	GR00T	0.44	1.4

Mean score by lighting:

	very bright	bright	dim
MolmoAct2	24.1	25.3	14.3
π0.5	19.0	17.8	8.7
GR00T	3.3	1.0	0.0

Most points come from reaching. Here is the furthest stage each model reached across its 60 rollouts:

	nothing	reach	grasp or later
MolmoAct2	25	29	6
π0.5	28	30	2
GR00T	57	3	0

Layout

meta/ data/ videos/observation.images.*   merged LeRobot v3.0 dataset, 180 episodes, 30 fps
results.jsonl                             one line per comparison (scores, preference, episode indices)
clips/<session>/<comparison>_{A,B}.mp4    standalone third-person clip of every rollout (H.264)
sessions/<session>/                       the same rollouts as one LeRobot dataset per scene session (pre-merge)

Episode features:

• observation.state: 6 joint positions in degrees, LeRobot v3 SO-101 calibration (arm frame)
• action: 6 joint targets as commanded to the arm
• observation.images.front: RealSense D435, third-person view, 640×480
• observation.images.wrist: wrist camera, 640×480

The models were trained in v2.1 units. To convert from the arm frame to model units, use model = signs * arm + offsets, with signs = [1,-1,1,1,1,1] and offsets = [0,90,90,0,0,0] (the gripper is unchanged).

from lerobot.datasets.lerobot_dataset import LeRobotDataset
import json, huggingface_hub as hub

ds = LeRobotDataset("ishneet/so101-pairwise-eval")
rows = [json.loads(l) for l in open(hub.hf_hub_download("ishneet/so101-pairwise-eval", "results.jsonl", repo_type="dataset"))]
r = rows[0]
ep = r["A"]["merged_episode_index"]   # rollout of model A in comparison r

Each results.jsonl row has these fields:

• comparison_id, scene (setup | lighting), task_id, task, prompt
• A / B, each with: model, score (0–100), note, episode_index (within its session), merged_episode_index, video, seconds, chunks, stopped_by, median_round_trip_ms
• preference (a model name or tie), preference_note

Protocol

• Tasks. Each task has 3–5 stages, and a rollout scores the furthest stage it reached:
  1. pick pen
  2. pen into glass
  3. pick highlighter cap
  4. pick highlighter
  5. pick bit
  6. bit into empty hole
  7. pick paper ball
  8. write S with pen
  9. kettle start
  10. kettle open
• Scenes. The table setups were normal table, white mat, green mat, cluttered, cluttered + green mat and cluttered + white mat. Each was run at three lighting levels: very bright, bright and dim.
• Blind A/B. A and B are run back-to-back from the same object placement, and the operator scores each rollout without knowing which model is which. Models are revealed after the scene.
• Execution. All models start from the same ready pose. Each predicted action chunk is executed open-loop at 30 Hz before the next query (as in the MolmoAct2 paper). Per-tick joint moves are capped at 15° for safety. A rollout lasts at most 60 s, or the operator stops it once it clearly fails or succeeds.
• Inference. Models ran on a remote A6000; the median round-trip was about 750 ms for MolmoAct2 and π0.5 and about 360 ms for GR00T. Each model sees both cameras at its own native input size:
  • MolmoAct2: 378², squash, 30-step chunk
  • π0.5: 224² letterbox, padded camera slots, 50-step chunk
  • GR00T: 256² squash, 40-step chunk

Caveats

• GR00T. The exact processor revision the GR00T checkpoint pins is not public, so we used upstream LeRobot. As a check, we compared each model's actions with the recorded actions in its own training data. GR00T's base turns in the right direction, but it barely beats holding still (mean error 5.65° vs MolmoAct2 1.85° on battery_bin_center). The checkpoint authors confirmed the setup is correct.
• Dim lighting. The wrist camera is nearly black in the dim scenes.
• Setup and pair are confounded. Each scene used a single pair, so a setup's per-model averages come from only one pair. Read setup breakdowns with care. Lighting and task breakdowns are balanced.
• Relabeled scenes. Two early scenes ("normal table", "white mat") were recorded before the lighting axis existed. They were relabeled "very bright", which matches the conditions they were recorded in.
• Scale. One arm, one operator, one environment.