Claude Code agent cracks Push-T robotics with 100% success, 46% fewer steps
Claude Code's Fable 5 agent masters Push-T with zero demonstrations, beating diffusion policies
A new arXiv paper from UC Berkeley researchers Shuangyu Xie, Kaiyuan Chen, and Ken Goldberg revisits the iconic Push-T manipulation task through the lens of agentic robotics. In the benchmark, a robot must use a single point of contact to push a T-shaped block into a target pose. Traditionally, this requires learning visuomotor policies from hundreds of human demonstrations. The team instead prompted an LLM coding agent — Claude Code with Fable 5 — to create an algorithmic solution without any demonstration data. The agent independently discovered a 2D gym simulation of the task, ran simulation experiments to learn push mechanics, and iteratively optimized its code until it achieved a 100% success rate.
Remarkably, the resulting code-as-policy used 46% fewer steps than the best diffusion policy trained on 200 human demonstrations. The agent also went beyond T: it solved the task for the full alphabet (Push-A through Push-Z) using a self-generated curriculum, and it produced simulation code for Franka and UR5 robot arms in 3D cross-embodiment scenarios with visual feedback. This demonstrates a significant shift — instead of collecting and training on human data, an agentic coding loop can autonomously discover, simulate, and solve robotic control tasks. The team plans to release videos, policies, and additional details online.
- Claude Code with Fable 5 achieved 100% success on Push-T using zero human demonstrations
- Agent's policy used 46% fewer steps than diffusion policy trained on 200 demos
- Generalized from Push-T to Push-A through Push-Z with self-generated curriculum and 3D cross-embodiment simulations for Franka and UR5 arms
Why It Matters
LLM coding agents could replace human demonstration data in robotics, dramatically cutting training costs and enabling autonomous skill discovery.