How bai-zihan/robotwin works
RoboTwin is a bimanual manipulation simulation benchmark and configurable trajectory generator, with a broad set of task environments and a collection of robot-learning policies. The documented user workflow is to select a task and configuration, run the task to collect trajectories, then process data for policy training and evaluation/deployment. The task base source supports simulation, camera observations, and robot motion planning. Policy families and processing capabilities are represented from the repository tree; their internal wiring was not established by the sampled excerpts.
Main entry points: train.py Policy Training · _base_task.py Task Environment · record_sim_episodes.py Episode Collection · envs Manipulation Tasks
Components 20
Task Simulation
- Manipulation Tasks
envs - Task Environment
envs/_base_task.py - Robot Control
envs/robot - Camera Sensing
envs/camera/camera.py - Scene Randomization
envs/utils/rand_create_actor.py
Trajectory Data
- Episode Collection
policy/ACT/record_sim_episodes.py - Recorded Trajectories
- Data Processing
policy/DP/process_data.py - Task Descriptions
code_gen/task_info.py
Policy Models
- ACT Policy
policy/ACT/act_policy.py - Diffusion Policy
policy/DP/diffusion_policy - DP3 Policy
policy/DP3/3D-Diffusion-Policy/diffusion_policy_3d/policy/dp3.py - Vision-Language Policies
policy/DexVLA/dex_vla/models/modeling_dex_vla.py - Additional Policies
policy - Policy Training
policy/RDT/train/train.py
Robot Execution
- Policy Deployment
policy/ACT/deploy_policy.py - Robot Task Runtime
envs/_base_task.py - OpenPI WebSocket Client
policy/pi0/packages/openpi-client/src/openpi_client/websocket_client_policy.py
Other components
- Benchmark User
- Inference Server
Connections 20
- Benchmark User → Manipulation Tasks: selects task
- Manipulation Tasks → Task Environment: defines environment
- Task Environment → Scene Randomization: creates scene actors
- Task Environment → Robot Control: plans motion
- Task Environment → Camera Sensing: captures observations
- Benchmark User → Episode Collection: starts collection
- Episode Collection → Task Environment: runs episodes
- Episode Collection → Recorded Trajectories: writes trajectories
- Recorded Trajectories → Data Processing: loads trajectories
- Data Processing → Policy Training: prepares training data
- Policy Training → ACT Policy: trains
- Policy Training → Diffusion Policy: trains
- Policy Training → DP3 Policy: trains
- Policy Training → Vision-Language Policies: trains
- Policy Training → Additional Policies: trains
- Benchmark User → Policy Deployment: starts evaluation
- Policy Deployment → Robot Task Runtime: runs policy
- OpenPI WebSocket Client → Inference Server: sends observations
- Inference Server → OpenPI WebSocket Client: returns actions
- Task Descriptions → Manipulation Tasks: describes tasks