Ryukijano/robojepa-rankability

World-model action rankability

A world model can plan by imagining the result of each candidate action and picking the one whose imagined future is closest to a goal image. This dataset tests whether those picks match what really happens. It contains fixed robot states ("anchors") from RoboCasa kitchens, 64 fixed candidate actions per state, and the true result of every candidate, measured by replaying it in the MuJoCo simulator. It also contains the scores of all eight released RoboJEPA models (22M to 8B parameters) on every candidate.

The Space has the leaderboard, a state explorer, and an in-browser scorer for your own model's scores.

Files

Config	Rows	Contents
anchors	3,000 pretrain + 499 target	Start and goal frames from the agent and wrist cameras (320×240 PNG), 7-D end-effector pose, goal position, base rotation, start distance to the goal d0_m, replay determinism, and a stuck flag
candidates	64 per anchor	The 7-D action (end-effector translation in the mobile-base frame, then 4 zeros), how it was generated (random, goal_coarse or goal_fine), the final distance to the goal after replay, the reward, and the final end-effector position
scores	one row per checkpoint, condition, anchor and candidate	RoboJEPA's planning score score_neg_l1 (minus the mean absolute difference between the predicted next features and the goal frame's features), a cosine score, and the imagination error (the same difference between the predicted and the real next frame, for the first 4 candidates)

leaderboard.csv has the metrics for every scored checkpoint. leaderboard_trained.csv has the models we trained or fine-tuned, with per-seed means, standard deviations and run identifiers.

The pretrain split covers the 30 RoboCasa pretraining kitchens with 100 seeds each. The target split covers the 10 held-out RoboCasa target kitchens with 50 seeds each, to test generalisation to kitchens the models never saw.

Scored models: all eight released RoboJEPA predictor sizes at their final checkpoint on both splits, every released checkpoint on target, and the 22M and 1B models with Gaussian pixel noise (σ = 25 on a 0 to 255 scale) added to the start frames. The compute axis is the released formula C = pre·1000·k + cool·5000 FLOPs.

How the ground truth is measured

The task is RoboCasa Reach with the PandaOmron robot and the robot0_agentview_left and robot0_eye_in_hand cameras. Each candidate is a 4-step action chunk (one world-model step) with a reach of at most 0.18 m. Every anchor has the same deterministic candidate bank for every model: 32 random actions within 0.30 m, 16 coarse and 16 fine goal-directed actions. We replay each candidate from a saved MuJoCo state; replaying the same action twice gives the same end position to within 0.2 mm. RoboJEPA runs with the V-JEPA 2.1 ViT-G/384 encoder, embodiment 2, one frame of context and fp16.

Metrics

Each metric is computed per anchor and then averaged.

• Overall Spearman: rank correlation between the score and minus the final distance, over all 64 candidates.
• Fine Spearman: the same over the 32 goal-directed candidates, whose outcomes are close together.
• AUROC: how well the score separates candidates that end within 5 cm of the goal.
• Regret: final distance of the top-scored candidate minus that of the best candidate.
• Hit@1 and Hit@3: whether the best candidate is the top 1 or in the top 3 by score.
• Pairwise accuracy: fraction of candidate pairs ordered correctly, with score ties counted as 0.5.
• ECE: calibration error of a logistic fit from the z-scored score to P(final distance < 5 cm), fitted on even anchors and measured with 10 bins on odd anchors, in sorted anchor_id order.

Reference code: rankability_metrics.py (Python) and metrics.js (the Space's in-browser scorer). Use final_dist_m at full precision; rounding it creates ties on anchors where the arm barely moves and inflates Spearman by up to about .02. ECE depends on the even/odd split, so it can differ by up to about .006 from our companion report, which split in collection order. All ranking metrics match the report exactly.

Results

Ranking improves from 22M to about 300M to 1B parameters (overall Spearman .747 to .789 on pretraining kitchens, .758 to .799 on held-out kitchens) and then stays within ±.002 up to 8B. Imagination error keeps falling over the same range (.273 to .218). Among the goal-directed candidates, the best models reach a fine Spearman of about .42, order pairs correctly about 65% of the time, and pick the single best candidate about 5% of the time.

Models we trained or fine-tuned

The predictors are trained from scratch on 500 RoboCasa365 simulator demonstrations. The ranking fine-tunes start from RoboJEPA-22M at 150k steps and train on 20 anchors per kitchen from pretraining kitchens 1 to 25, using replayed outcomes. All numbers below are on the 499 target anchors from the 10 held-out kitchens, none of which are used for training.

Model (id)	Type	Image encoder	Parameters	Training data, steps	Seeds	Overall ρ	Fine ρ
RoboJEPA-22M, released (robojepa-22m-150k)	released	V-JEPA 2.1 ViT-G	22M	none (frozen)	1	.758	.357
RoboJEPA-1B, released (robojepa-1b-300k)	released	V-JEPA 2.1 ViT-G	1B	none (frozen)	1	.799	.426
Deterministic predictor (det_sub500_long)	trained	V-JEPA 2.1 ViT-G	18.8M	500 demos, 30k	1	.736	.298
Deterministic predictor (det_sub500)	trained	V-JEPA 2.1 ViT-G	18.8M	500 demos, 6k	1	.713	.308
Flow matching, velocity target (flow_sub500_long)	trained	V-JEPA 2.1 ViT-G	18.8M	500 demos, 30k	1	.606	.129
Flow matching, clean-frame target (flowx_sub500_long)	trained	V-JEPA 2.1 ViT-G	18.8M	500 demos, 30k	1	.622	.151
Flow matching, velocity target (flow_sub500_90k)	trained	V-JEPA 2.1 ViT-G	18.8M	500 demos, 90k	1	.723	.254
Deterministic predictor (det_dinol_sub500_long)	trained	DINOv3 ViT-L	18.8M	500 demos, 30k	1	.659	.215
RoboJEPA-22M + ranking loss (cooldown_22m_rank_pilot)	fine-tuned	V-JEPA 2.1 ViT-G	22M	kitchens 1 to 25, 20 anchors each, 1k	3	.823	.492
RoboJEPA-22M, same fine-tune without ranking loss (cooldown_22m_control_pilot)	fine-tuned	V-JEPA 2.1 ViT-G	22M	kitchens 1 to 25, 20 anchors each, 1k	3	.794	.407
Random scores (random)	baseline	none	none	none	5	−.003 ± .004	.002 ± .008

Score your model

Write a parquet file with anchor_id, cand_idx and score columns, where a higher score means your model expects a better outcome. Each anchor you include needs a score for all 64 of its candidates. Then run the evaluator:

python evaluate.py --split target --scores my_scores.parquet --name my-model --out results/my-model.json

It loads the matching anchors and outcomes from this repository (or from a local copy with --data_dir data). To add a leaderboard row, open a pull request on this dataset with results/<name>.json and a short model description. We add the row once we can reproduce the result.

Limits and prior work

Ranking fixed candidate actions with a world model has been done before (ARC-Bench, AD-WM). This release adds a full released model-size ladder with its compute axis, held-out kitchens, held-out calibration, and imagination error measured on the same inputs as ranking. It covers one task family (Reach, one step ahead), one simulator and one scoring objective. The ranking fine-tune trains on anchors and candidates built the same way as the test set, so its gain on held-out kitchens shows transfer across kitchens, not across tasks. Everything is simulated; no physical robot is involved.

Licence

CC BY-NC-SA 4.0, because the scores are derived from RoboJEPA weights (CC BY-NC-SA 4.0). RoboCasa is MIT. Non-commercial research use only.