On July 9, 2026, Anthropic published "Claude plays robotics" on its official research blog, sharing the results of an experiment where Claude and other language models were given control over real and simulated robots. Written by researchers Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman, the study set out to answer a basic question: can a language model perceive a scene, understand a particular robot's state, and issue actions that reliably change the physical world? You don't need any robotics or programming background to follow along — here's what the test covered and what it found, in plain language.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What does it actually mean for Claude to "control" a robot?
In this study, "control" doesn't mean one single method — it covers four different ways a language model can be connected to a robot, each requiring the model to handle a different amount of low-level detail.
| Method | Level of abstraction | What Claude actually does |
|---|---|---|
| Directly commanding motor torque (motor torque — the amount of force applied at a robot joint) | Lowest | Calculates the exact force to apply at each joint |
| Writing controller code | Low | Writes the program that moves the robot |
| Training a controller with reinforcement learning (RL — a method where the model learns by trial and error to maximize a reward) | Medium | Trains a controller from scratch through trial and error |
| Giving high-level instructions to a pretrained robot policy (pretrained robot policy — a separate neural network that turns high-level commands into coordinated joint movements) | Highest | Issues broad instructions like "walk forward," leaving fine motor detail to an already-trained separate program |
Note: This was a controlled research experiment run by Anthropic's research team, not a feature you can access directly through claude.ai or the Claude app. Keep in mind this is a robotics research result, not a consumer-facing capability.
Which robots and tasks were used in the test?
The researchers tested language models across a range of robot bodies — from classic control toys to a humanoid robot to a real legged robot — and across three categories of tasks. For real-world testing, they used a Unitree Go2 quadruped robot, which the blog post notes is the same robot used in the earlier Project Fetch.
| Task category | Examples | Robot bodies used |
|---|---|---|
| Classic control | Classic reinforcement-learning problems like balancing an inverted pendulum or controlling a hopper (a single-legged jumping robot) | Classic control toy robots |
| Locomotion and navigation | Getting legged robots to balance, walk, and move through space | Simulated quadruped and humanoid robots, real Unitree Go2 |
| Manipulation | Using a robotic arm to grasp and move objects | Robotic arm |
What were the key findings?
The central finding is that Claude's robotics performance depends as much on how it is connected to the robot as on the model itself. When forced to control joints directly, models mostly failed — but when supervising a pretrained controller or using simple orientation tools, they could complete real navigation and manipulation tasks. Newer models were also notably better at adjusting their strategies and converting image and sensory understanding into appropriate actions.
- Improvement across model generations is uneven: high-level interfaces (using a pretrained policy) improve consistently with each new model, while direct low-level control improves inconsistently — some models get clearly better, others don't.
- Even today's frontier models cannot control a humanoid robot on their own without a pretrained policy.
- Still, the researchers reported real gains in both direct manipulation and high-level policy control across the humanoid and quadruped bodies tested.
How did locomotion and manipulation tasks perform specifically?
On locomotion, frontier models can now perform limited but meaningful whole-body control. There were clear gains in low-level control for quadruped standing, balancing, and walking, along with weaker but measurable gains in humanoid balancing. Combined with pretrained policies and perception tools, models could even navigate simple environments. However, they still failed at tasks requiring stable spatial memory, self-localization, or long open-loop plans.
On manipulation, low-level direct control showed clear improvement in local physical behavior — reaching objects, making contact, and grasping — but full task completion remained rare, ranging from 0 to 5.5%. With high-level methods, using VLA (vision-language-action — a pretrained policy that maps camera images and an instruction directly to robot-arm motions) scaffolds raised manipulation performance far above direct control. Newer models were also better supervisors of these policies: better at recognizing when a proposed action would fail, and less likely to defer to the policy indiscriminately. Supervising a policy still cost some performance compared to running the policy alone, but the best supervisors closed most of that gap — and on tasks the policy couldn't do alone, the strongest models' supervision provided a net improvement.
Why does this research matter?
Anthropic ties these findings directly to the challenge of safely deploying language models in the physical world. The fact that today's frontier models cannot control humanoid robots without a pretrained policy is an important current limitation, but the researchers also emphasized that real gains are appearing across model generations in both direct manipulation and high-level policy control. As a concrete example, they noted that a general-purpose chat model with no robotics-specific training can, on a good run, write and download its own tools to slowly walk a quadruped robot through a maze, or pick a plate off a counter and set it on a stove — and that the reliability gap for tasks like these is closing with each model generation.
Frequently Asked Questions
Q. Can I control a robot directly through claude.ai or the Claude app right now?
No, this was a research experiment run internally by Anthropic's research team, not a consumer feature available through claude.ai or the Claude app.
Q. Is Claude actually good at controlling robots?
It depends on the connection method. Giving Claude direct control over fine details like joint force mostly fails, but letting it issue high-level instructions to an already-trained robot control policy allowed it to complete real navigation and manipulation tasks.
Q. Was this tested on real robots, or only in simulation?
Both. Alongside simulated quadruped and humanoid robots, the official blog post states the team also tested a real Unitree Go2 quadruped robot.
Q. What's the name of the benchmark used in this study?
The researchers used a benchmark suite called Embody, which they used to measure composite scores across robot bodies and tasks.
This research is less about a feature that lets Claude move a robot for you, and more a look at how language models interact with the physical world and what that means for safety. If you're interested in other official Claude news and safety research, check out our other research roundups on usingclaude.com.
Related articles
- Claude web search & Research — using live search and deep research well
- Claude Projects, Artifacts & Memory: What They Are and How to Use Them
- How to Write Prompts for Claude: 6 Core Principles from the Official Docs
- How to Use Claude Projects: Knowledge & Instructions to Cut Repetition
- Claude Artifacts: How to Create, Edit, Publish & Share (2026)
- Claude Found Mathematical Weaknesses in Cryptographic Algorithms: What It Means