Fable 5.2 Tops RoboColiseum's New Season at 18.30, Then Scores 3.73 on Dexterous Hands
Claude Fable 5.2 leads RoboColiseum's new season at 18.30 overall and 33.04 on fine manipulation, but scores just 3.73 on dexterous hands. Policy models beat every agent there.
5 min read

Claude Fable 5.2 has taken first place in RoboColiseum's new season with a composite score of 18.30. It also leads on fine manipulation (33.04) and cross-embodiment transfer (17.99). The one category where the overall champion falls apart is the one that most resembles real work: dexterous hands, where it scores 3.73.
Hanging clothes on a rack, or holding an object steady without crushing or dropping it, remains unsolved for these models. The new season made that gap measurable.
Four New Tracks, 50 Tasks
RoboColiseum rebuilt its evaluation around tasks a robot would actually be asked to do: charging a phone, hanging clothes with a hanger, picking up trash, opening a door with a key. All of them require fingers, both arms, and body coordination to work together. The season adds four tracks and 50 tasks, scored both on whether a task is completed end to end and on how efficiently it is executed.
Dexterous Manipulation (10 tasks)
Hanging a shirt means grasping the garment, spreading it open, threading the hanger through, then adjusting and lifting it onto the rod. The fabric deforms and sags, multiple fingers must coordinate the grip, and both hands have to stay in sync through the sequence. The track also includes wok tossing, solving a Rubik's cube, tying a garbage bag, and carrying a package with both hands.
Full-Body Coordination (8 tasks)
Picking up trash requires moving to the right position, adjusting the torso, then reaching for the object. Erasing a whiteboard requires driving the base and tilting the torso so the arm can reach the full surface. The track covers trash pickup, placing plates, retrieving items from a conveyor, unlocking a door with a key, and whiteboard erasing—testing whether a model can chain locomotion and manipulation without breaking the task.
Fine Manipulation (8 tasks)
Charging a phone means two arms independently holding the device and the connector while aligning position and orientation at the same time. The track covers phone charging, assembling blocks, plugging in an Ethernet cable, mounting a bearing, and packaging a phone—precision work where accuracy requirements rise as the task approaches completion.
Cross-Embodiment Transfer (6 task types, 24 tasks)
Hanging a cup, swapping keycaps, writing with a pen, pouring water—each task type is tested on four different robot bodies (ARX, AgileX, G1, G2), producing 24 evaluation tasks. Joint structure, range of motion, and gripper change, so the model must adapt its actions accordingly. This track measures whether one model can serve multiple platforms.
The Full Leaderboard
The current board counts 8 agents and 3 policy models.
| Model | Composite | Fine manipulation | Cross-embodiment |
|---|---|---|---|
| Claude Fable 5.2 | 18.30 | 33.04 | 17.99 |
| GPT-6 Astra | 13.73 | — | — |
| Claude Opus 5.5 | 11.77 | — | — |
| π0.5 (policy) | 11.22 | — | — |
In full-body coordination, Astra scores 18.96 and Opus 18.94—effectively a tie, and both close to the category ceiling.
Dexterous Hands: Everyone Fails
All eight agents hit their personal lowest score on the dexterous-hand track.
| Model | Dexterous hand score | Type |
|---|---|---|
| π0.5 | 16.79 | policy |
| LingBot-VLA-2.0 | 10.66 | policy |
| GR00T N1.7 | 5.53 | policy |
| Gemini 4 Argon | 5.19 | agent |
| Claude Fable 5.2 | 3.73 | agent |
| GPT-6 Astra | 0.00 | agent |
| GPT-6.1 Terra | 0.00 | agent |
Gemini 4 Argon's 5.19 is the best any agent managed. Fable 5.2, first overall, drops to 3.73. Astra and GPT-6.1 Terra score zero.
The pattern inverts for policy models: all three beat all eight agents on this track. π0.5's 16.79 is more than four times Fable 5.2's score. Coordinating multiple fingers into a stable grasp and then manipulating the held object is the exact capability these general models were not trained for, and purpose-built vision-language-action policies are.
Three Anti-Gaming Mechanisms
Repeatedly practicing the same task set lets a model memorize rather than generalize. The season adds three checks to suppress inflated scores from task memorization and privileged information.
Hidden Test Set
Open scenes are available for development and debugging. The official test scenes on the server follow the same distribution but use different layouts, so models are evaluated on arrangements they have not trained on.
Dynamic Sampling
A fixed official set produces the reported score. Alongside it, a dynamic set changes with every submission. It does not count toward the total; instead it is compared against the fixed set. If the gap between the two is too large, the submission is restricted from the leaderboard.
Human Review
Beyond automated checks, reviewers screen for behavior that violates evaluation rules to obtain scores improperly.
Agents and Policies Ranked Separately
Agents can alternate between observation, reasoning, and action. Vision-language-action and world-action models typically execute a fixed action sequence. The two interaction patterns differ in rhythm, so the season separates them into distinct leaderboards for like-for-like comparison.
Execution Efficiency Enters the Score
Agents now face a limit on execution steps, with the threshold derived from training data. A model that repeatedly observes, thinks, and retries past a reasonable count is penalized according to how far it overshoots. Partial progress still earns credit, and finishing within the prescribed step budget earns a bonus. How far a run gets, whether it completes, and how many steps it took all affect the final score.
Proving the Simulator Is Trustworthy
A simulation is only worth what its correlation with reality. RoboColiseum ran a large-scale control experiment: no adaptation or tuning of the simulator, models trained in the real world dropped straight into simulation for zero-shot real-to-sim evaluation.
Across 4 robot bodies and 22 complex tasks, the correlation between simulation scores and real-robot task scores reached 96.86%. Simulator scores are therefore not a separate virtual report card—failures found in simulation and strategies validated there transfer to hardware with high confidence, which cuts the trial-and-error cost of real-robot debugging.
Submit a Model Through a Conversation
The pipeline—prepare data, connect a model, submit for evaluation, read results—runs through dialogue with an AI. RoboColiseum exposes AI-invocable skills, so after installation and configuration a developer can ask the AI to download training data, connect an inference service, and launch an evaluation. Once a job is created and inference starts, the platform handles simulation runs, scene scheduling, and scoring, and exposes progress, per-category scores, and results. The intent is to return engineering time to failure analysis and model improvement rather than harness setup.
What the Season Shows
General-purpose agents now handle precision assembly and body coordination far better than multi-finger manipulation. The three categories where scores remain low—dexterous control, fine contact work, and cross-embodiment transfer—are the same three where specialized policies already outperform the generalists. For teams working on dexterous control, that gap is the opening: the frontier models have not solved it, and the policy stack has a measurable head start.



