Anthropic Claude Robotics Evaluation Shows Rapid Gains on High‑Level Control but Limited Direct Torque Mastery

TL;DR

Anthropic’s July 2026 robotics benchmark shows that the latest Claude models can successfully guide pretrained locomotion and manipulation policies to complete real‑world tasks, but they still fail when asked to issue raw motor torques or forces directly.

Key Finding: Control Interface Determines Performance

Interface matters as much as model size. The same Claude model can score near‑zero when issuing raw torques, yet achieve the highest composite score on the Embody benchmark when supervising a pretrained policy.
High‑level interfaces (policy supervision, VLA scaffolds) consistently outperform low‑level ones. Direct torque control fails for all models on humanoid and quadruped bodies; programmatic or policy‑based control yields measurable success.

Generational Improvements Are Uneven

High‑level gains are clear. Claude Opus 4.5 → Opus 4.7 and Mythos Preview show large jumps in the high‑level locomotion composite (up to +15 points).
Low‑level control improves only modestly. Direct‑control scores rise slightly across generations, but many newer models (e.g., Opus 4.6) still cannot stand a quadruped from a prone pose.
Reinforcement‑learning supervision lags behind code control. Only GPT‑5.4 reliably learns a competent policy for the chaotic TwinFlipper task; other models perform worse than when they simply write Python controllers.

Locomotion Results

Robot	Task	Best‑performing Claude model	Interface
Unitree Go2 (quadruped)	Balance (≈2 s)	Opus 4.6 / Mythos Preview	Programmatic (Python)
Unitree Go2	Walk forward	Opus 4.6 (programmatic)	—
Unitree G1 (humanoid)	Stand from collapse	None (no model succeeded)	Direct / Programmatic

High‑level navigation (11 tasks, e.g., find_x, maze, drift_detection) reaches a composite of 54/100 for Mythos Preview using a pretrained joystick policy and a compass tool.
Perceptual aids: Adding a compass consistently lifts performance across models; a third‑person camera helps only the strongest models (Opus 4.7, Mythos Preview).

Manipulation Results

Platform	Task	Best Claude model	Success rate
Franka Panda arm (LIBERO)	Direct end‑effector control	Mythos Preview	5.5 % overall task success
Same platform	VLA‑supervised manipulation (40‑task LIBERO)	Opus 4.6 / Opus 4.5	~30 % task success (vs. 70 % for VLA alone)

Direct control: Models increasingly reach the target object and make contact, but full grasps remain rare (0–5 %).
VLA supervision: All models improve when they can accept, edit, or replace VLA actions. Claude models follow VLA recommendations more often than GPT‑5.4 or Gemini 3.1, reducing the “control penalty.”
Novel tasks: Opus 4.5‑4.7 recognize VLA failures and defer less, achieving higher net uplift than older models that blindly follow the policy.

Vision and Tool Ablations

Depth maps, segmentation, and crosshair provide little benefit for manipulation; the cursor tool (interactive point‑query) raises Mythos Preview’s success from 6 % to 32 % on a 10‑task subset.
Compass is the most effective aid for locomotion, improving every configuration by 5–12 points.
Raw images vs. textual scene description: Older Claude models perform better with text descriptions, indicating weak pixel‑level spatial reasoning. Newer Claude models (Opus 4.6, Opus 4.7) and Gemini lose performance when images are replaced, showing they already extract useful visual information.

Reasoning Budget Effects

Minimal impact on classic control, locomotion, and manipulation for most Claude generations.
Exceptions: Mythos Preview gains noticeably from an adaptive‑max reasoning budget; GPT‑5.4 benefits on locomotion; Gemini 3.1 shows mixed effects.
Overall: Extra reasoning alone does not close the gap between low‑level and high‑level control.

Short‑Term In‑Context Learning

Classic control: Later Claude models improve primarily by iterating over failed attempts, not by starting stronger.
Context truncation: Removing early turns rarely harms performance; sometimes it helps ("context rot"). Mythos Preview remains robust, suggesting it has baked‑in strategies.
High‑level locomotion practice runs: Mythos Preview learns a simple L‑shaped course after a single example, whereas Opus models require several repetitions.

Real‑World Validation on a Physical Unitree Go2

Find‑X task: Models often stop short of the required 1 m proximity; older models mis‑align or mistake reflections for the target.
Office‑loop navigation: All models fail due to vision‑memory errors (missing turn cues, overshooting corridors).
Tool usage: Crosshair helps the model judge alignment but can mislead it about obstacles (e.g., walking into a trash can that appeared left of the crosshair).

Safety Implications

Capability swings: Small changes in tooling (e.g., adding a compass) can increase a model’s physical influence by orders of magnitude.
Evaluation must include access level: Benchmarks that only test isolated LLMs under‑estimate real‑world power when the model is coupled with a competent controller.
Supervision vs. autonomy: The best current approach is to let a language model supervise a pretrained policy rather than replace it; this limits unsafe low‑level actions while still providing useful guidance.

Takeaways for Researchers and Practitioners

Prioritize high‑level interfaces when deploying LLMs on robots; direct torque control remains unreliable.
Provide orientation tools (compass, cursor) to mitigate perceptual bottlenecks.
Expect rapid generational gains in perception and short‑term adaptation, but do not assume long‑term planning abilities have arrived.
Design safety guards around policy supervision—allow the LLM to veto or adjust actions, but keep the low‑level controller as the final arbiter.
Benchmark with the full stack (LLM + policy + tools) to obtain realistic capability estimates.

All figures, tables, and quantitative scores are reproduced from Anthropic’s original post “How Claude Performs on Robotics Tasks” (2026‑07‑09). No additional data were invented.