RoboAgentLab

Agents that Act, Build, and Evolve.

An open research platform for agentic robotics.

Website · Get started · Platform code · Awesome list · Datasets

Robotics as software engineering

RoboAgentLab brings robot task execution and experimental development into a shared coding-agent workspace. Built on pi, the platform combines an agent loop and persistent sessions with robotics extensions. Downstream servers provide simulation, robot control, perception, planning, and policy inference; upstream skills explain how to compose these capabilities. The same interfaces are available to agent tool calls and generated Python programs.

Three complementary roles

Role	Research focus
Act — Agent as Policy	Select behaviors, orchestrate tools, and write programs that execute robot tasks.
Build — Agent as Infrastructure	Develop and debug robotics systems, reproduce experiments, and connect data collection, training, and evaluation.
Evolve — Agent as Self-Evolver	Use feedback to revise and retain reusable components, workflows, and research capabilities.

Two streams of research data

Robot policy interactions — Agent interactions with robot environments, including observations, decisions, tool calls, and feedback. Our research goal is to distill large-model teachers into compact policy agents that combine task accuracy with fast inference.
Robotics R&D interactions — Coding-agent interactions from robotics development, integration, debugging, and testing. Our research goal is to use these traces for domain-specific post-training of open-weight models that can drive the RoboAgentLab harness.

The dataset cards describe the contents and structure of the current releases.

Research and open resources

Our forthcoming paper, Agents that Act, Build, and Evolve: A Unified Framework and Platform for Agentic Robotics, connects these roles within a common framework. The RoboAgentLab platform and Awesome Agentic Robotics are available now.

Explore the platform repository and the datasets above to follow the project.