Platform codeMAKE YOUR NEXT EXPERIMENT REUSABLE
Collection
Create worlds and experience
OPEN RESEARCH PLATFORM
Agents that
act, build,
and evolve.
One workspace for agentic robotics.
Connect models, tools, and robots.
Turn ideas into experiments.
FROM CODE TO PHYSICAL INTERACTION
A world of experiments.
Manipulation, assembly, and everyday tasks.
Explore a selection of environment replays.
A SHARED HARNESS. OPEN ENDED RESEARCH.
Build once.
Compose across robots.
pi supplies the agent loop and sessions. Robotics capabilities connect through extensions, run as downstream services, and come with skills that explain how to use them.
↓The agent selects a registered toolObservations & results ↑
↓Dispatch an operation over HTTPImages · state · outcomes ↑
DOWNSTREAM · EXECUTABLE CAPABILITIESLocal or remote services
Service results return through the extension to the next agent turn.
A FRAMEWORK FOR AGENTIC ROBOTICS
One framework.
Three agent roles.
Select a category to explore papers and code.
01 / ACT
Agent as Policy
Observation inputs × execution authority
| EXECUTION → OBSERVATION ↓ | A1 Behavior selection Fixed executor | A2 Tool orchestration Compose tools | A3 Code synthesis Robot programs |
|---|---|---|---|
|
O1
Sensory input
Images, depth, touch
|
Code Browse ↗ Model Browse ↗ | Tool calls Browse ↗ | Programs Browse ↗ |
|
O2
Perception tools
Outputs and queries
|
Code Browse ↗ Model Browse ↗ | Tool calls Browse ↗ | Programs Browse ↗ |
|
O3
Privileged state
Simulator state
|
Code Browse ↗ Model Browse ↗ | Tool calls Browse ↗ | Programs Browse ↗ |
↑Resources and tools
Execution feedback↓
02 / BUILD
Agent as Infrastructure
Collection–training–inference
Training
Run learning and experiments
Inference
Evaluate and diagnose execution
↑Retained revisions
Development feedback↓
03 / EVOLVE
Agent as Self-Evolver
Retained updates from feedback
Iterative artifact refinement
Component optimization
Harness improves harness
Works can appear in more than one category.
THE PAPER · FORTHCOMINGAgents that Act, Build, and Evolve:
Read the companion list ↗Agents that Act, Build, and Evolve:
A Unified Framework and Platform for Agentic Robotics
LEARN FROM BOTH SIDES OF THE EXPERIMENT
Interaction becomes
research data.
Two training routes: distill fast policy agents from robot interactions, and specialize open models through Robotics R&D interactions.
Explore each stage of the two training routes.
AGENT ↔ ROBOT
Distill a compact policy agent
Large-model decisions → fast robot task execution
AGENT ↔ DEVELOPMENT
Post-train a Robotics R&D model
General-model expertise → robotics-specialized development