Event
RAX Open-Source Framework: Coding Agents Learn Robot Tasks through Online Exploration
3 reports · 3 sources
Runhan Huang et al. (team from Harvard, NVIDIA, Berkeley, MIT, and Stanford) proposed and open-sourced RAX (Robot Agent eXploration), a robot agent framework: a general-purpose coding agent directly generates robot control code, without using action demonstration data, hand-designed action primitives, or value function guidance; model parameters remain frozen, and capabilities accumulate through online exploration. The paper claims that on five simulation benchmarks—RoboDojo, BEHAVIOR-1K, HumanoidArena, DexJoCo, and LIBERO-Pro—RAX leads the previous strongest baseline by an average of 37.4 percentage points; on real robots, with zero teleoperation demonstrations, it can perform tasks such as precision peg insertion, deformable objects, articulated objects, and long-horizon mobile whole-body control after less than 2 hours of autonomous exploration, and it continuously improves performance by composing skills validated in prior exploration. RAX is open-sourced as a general-purpose robot agent platform; new robot embodiments can be integrated solely through its API, without teleoperation data collection or policy training; the team also released the Agent instructions, workflow, robot interface, benchmarks and evaluation protocols, as well as the control programs and skills generated during exploration.
Report timeline
- Harvard, Nvidia and others propose RAX: robot agent online exploration, leading by 37.4 points on average across five benchmarks
A research team from Harvard, Nvidia, Berkeley, MIT and Stanford proposed RAX (Robot Agent eXploration), which connects a general coding agent to robot perception and control interfaces, uses online exploration to write successful decisions into executable programs and accumulate them as skills, while keeping model parameters frozen.
CVer · 微信公众号 ·
- Harvard, Berkeley, NVIDIA and others propose RAX, using online exploration scaling to let robots learn skills with zero demonstrations
A research team from Harvard, NVIDIA, Berkeley, MIT and Stanford proposes RAX (Robot Agent eXploration), which connects a general coding agent to robot perception and control interfaces, uses online exploration to write successful decisions into executable programs and accumulate them as reusable skills, while keeping model parameters frozen.
机器之心 · 微信公众号 ·
- RAX: An open-source framework for scaling robot agent capabilities through online exploration
The paper proposes RAX (Robot Agent eXploration), an open-source robot agent framework centered on scaling online exploration, in which a general-purpose coding agent generates robot control code without using action demonstration data, hand-engineered action primitives, or value-function guidance.
arXiv cs.RO ·