Beating Fable 5.2 with Old Claude: Q0-Preview's Agent Approach Tops RoboDojo

Background: Fable 5.2 Sets New Benchmark

On November 10, Anthropic released Fable 5.2, achieving 64.9% accuracy on the RoboDojo Sim dual-arm robot simulation benchmark, a significant jump over Fable 5.1 and a clear lead over GPT-6 Astra. The model demonstrated capabilities like unscrewing bottle caps, manipulating Rubik's cubes, and drawing.

Q0-Preview Surpasses Fable 5.2 with a Different Paradigm

As of November 11, Q0-Preview from Qoo Tech (乾章智能) tops the RoboDojo Sim Agent leaderboard with a 68.87% task success rate and a 74.2 composite score — the first entry to break 70. Unlike VLA (Vision-Language-Action) or WAM (World Action Model) approaches, Q0-Preview adopts an Agent architecture that exploits a frontier multimodal LLM's planning, spatial perception, memory, and tool-use abilities to autonomously execute complex physical tasks.

Surprisingly, the backbone model is not GPT-6 Astra but the previous-generation Claude Opus 5.5 , which Anthropic officially reported at only 39.4% accuracy. This raises a deeper question: where are the boundaries of general LLMs' physical action capabilities, and how can spatial reasoning and embodied know-how be better fused?

Core Idea: Spatial Computation as Embodied Operation

The central thesis: determine the contact location and contact pose for each manipulation stage. For a simple task like grasping an apple, after identifying the target object, the model must decide which part of the apple to grasp, from which direction the gripper should approach, and with what orientation. Converting these judgments into a target end-effector pose solves the problem.

Q0-Preview leverages the model's spatial perception and coding ability: using single/multi-view observations to locate target pixels, then applying camera intrinsics/extrinsics and pixel-to-3D tools to map image-space contact targets into world coordinates.

Unified representation: For a parallel gripper, three spatial reference points encode target pose and contact information — the midpoint of two ordered points defines the reference contact position, the line between them defines the grasp direction, and the third point adds the approach direction. Thus position, orientation, and contact become a computable spatial calculation problem.

The multimodal agent observes the scene and instruction, decides the current sub-task, and outputs a target pose plus gripper command. A downstream motion planner (e.g., cuRobo) converts this into continuous actions. After execution, the agent inspects the new observation, verifies the result, and adjusts the next step — forming an "observe → compute → execute → feedback" closed loop .

Harness: From "Can Calculate" to "Can Do"

Around the core idea, Q0-Preview builds a Harness that standardizes interaction between the model, observations, tools, and execution modules:

1. Explicit execution boundaries & tool calling: Unifies available observations, robot state, and prior knowledge; budgets thinking and tool calls; defines control output format. Tasks advance in short action chunks with visual checkpoints at critical intermediate stages. The model may call built-in or self-written programs for image cropping, local zoom, etc., to assist spatial computation.

2. Accumulating general manipulation experience via sim trial-and-error: During development, the team runs simulated trials to distill reusable operational heuristics — e.g., using the idle arm's wrist camera to reduce occlusion, planning paths that balance collision avoidance and future visibility, adjusting motion speed per grasp/transport/place phase. These skills transfer from sim to real.

3. Supplementing task specifications & completion conditions: For vaguely defined tasks, Q0-Preview adds static task rules, success criteria, and operational hints to reduce mismatch between model understanding and evaluation requirements. No real-time reward or environment judgment signals are read during execution.

4. Managing context & task memory: Retains initial info, recent observations, and a compact task memory recording confirmed progress, current state, and future goals, controlling context growth. Explicit interfaces isolate upstream decision-making from downstream execution, enabling seamless multi-stage chaining.

Demonstrated Capabilities

Dual-arm synchronous operation: Sets independent targets for left and right arms in the same phase, advancing parallelizable actions simultaneously.

Dual-arm coordinated operation: Allocates roles around a shared goal, coordinating execution based on dependencies like support and handover.

Precise manipulation: Uses multi-view spatial computation to determine position and orientation, then corrects deviations with close-up observations.

Adaptive adjustment: Re-localizes targets, revises poses, or re-plans based on object state changes and execution outcomes.

Memory: Retains previously acquired information and completed steps to inform later decisions, enabling tasks that depend on history.

Open Challenges

Inference efficiency & cost: Frontier LLMs may consume millions of tokens and hours of inference for seemingly simple embodied operations — far from practical deployment requirements. Reducing redundant reasoning and tool calls while preserving capability is a must-solve problem.

Source of embodied capability: GPT-6 Astra's debut sparked speculation that massive embodied ego-data drove its spatial and manipulation leap. Yet Q0-Preview's success with Opus 5.5 forces re-examination: which factors — embodied data, code generation, tool use, 3D generation — contribute most? How do model-inherent abilities and Harness design interact? Systematic ablation experiments are needed to guide future data collection, training objectives, and system design.

Latency vs. dynamic environments: Long reasoning latency limits responsiveness to dynamic changes. Real-world objects don't wait for the model to finish thinking; observation-to-execution drift can invalidate plans. Timely reaction, fusing vision with force/tactile feedback, and maintaining reliability over long horizons under disturbances remain open.

Outlook

Despite unsolved challenges, the zero-shot generalization of foundation models opens a new path for embodied intelligence. Q0-Preview shows that combining general model capabilities with domain knowledge and constraints can continuously expand robotic task boundaries. The team plans to release a technical report, faster and cleaner code, more real-robot experiments, and further discussion of capability limits and future directions.

Official evaluation code is open-sourced at:

https://github.com/XPolicyLab/XPolicyLab/tree/main/policy/Q0_Preview