How limxdynamics/tron2_openpi works
TRON2 OpenPI is a deployment-focused robot-policy stack: a robot client gathers observations, sends them to a policy server, and receives action chunks for execution. The README also describes a separate training path that produces checkpoints and normalization assets for deployment, plus RTC-assisted chunk execution and optional observation sources. Model internals are represented at the principal model boundary; the excerpts do not show all script wiring or the robot-client implementation, so uncertain internal links are omitted.
Main entry points: policy.py Robot policies · websocket_client_policy.py WebSocket policy client · tron2 RTC deployment client · websocket_policy_server.py WebSocket server
Components 22
Policy deployment
- Policy server launcher
scripts/serve_policy.py - WebSocket server
src/openpi/serving/websocket_policy_server.py
Policy processing
- Data transforms
src/openpi/transforms.py - TRON2 transforms
src/openpi/policies/tron2_policy.py - Policy construction
src/openpi/policies/policy_config.py - Robot policies
src/openpi/policies/policy.py - RTC guidance
src/openpi/rtc/rtc_processor.py
Model inference
- Pi0 model
src/openpi/models/pi0.py - Pi0 PyTorch
src/openpi/models_pytorch/pi0_pytorch.py
Training and assets
- Model checkpoint
- Training dataset
- Train policy
scripts/train_tron2_task.py - Training registrations
src/openpi/training/config.py - Data loading
src/openpi/training/data_loader.py
Robot clients
- WebSocket policy client
packages/openpi-client/src/openpi_client/websocket_client_policy.py - TRON2 deployment client
examples/tron2 - RTC deployment client
examples/tron2 - Observation sources
examples/tron2
Other components
- Task operator
- TRON2 robot
- TRON2 Bridge
- RealSense cameras
Connections 32
- Task operator → TRON2 deployment client: starts task
- Task operator → RTC deployment client: starts RTC task
- TRON2 deployment client → Observation sources: requests observations
- RTC deployment client → Observation sources: requests observations
- TRON2 Bridge → Observation sources: supplies observations
- RealSense cameras → Observation sources: supplies images
- Observation sources → TRON2 deployment client: provides input
- Observation sources → RTC deployment client: provides input
- TRON2 deployment client → WebSocket policy client: submits observations
- RTC deployment client → WebSocket policy client: submits RTC requests
- WebSocket policy client → WebSocket server: sends WebSocket request
- WebSocket server → Robot policies: dispatches inference
- Policy server launcher → Policy construction: constructs policy
- Policy server launcher → WebSocket server: starts serving
- Policy construction → Robot policies: creates policy
- Robot policies → TRON2 transforms: applies TRON2 transforms
- Robot policies → Data transforms: applies input transforms
- Robot policies → Pi0 model: runs inference
- Robot policies → Pi0 PyTorch: runs inference
- RTC deployment client → RTC guidance: configures guidance
- Pi0 model → RTC guidance: applies guidance
- Robot policies → WebSocket server: returns action chunk
- WebSocket server → WebSocket policy client: sends action chunk
- WebSocket policy client → TRON2 deployment client: returns actions
- WebSocket policy client → RTC deployment client: returns actions
- Training dataset → Data loading: provides samples
- Train policy → Training registrations: selects configuration
- Train policy → Data loading: loads training data
- Data loading → Data transforms: transforms samples
- Train policy → Model checkpoint: writes weights
- Model checkpoint → Policy server launcher: loads weights
- Training registrations → Policy construction: registers policy config