How robotwin-platform/robotwin works
RoboTwin is a bimanual robotic-manipulation benchmark and simulation platform. A user selects a task and policy; a task environment builds a SAPIEN scene, actors, cameras, and robot, emits observations, receives policy actions, plans and executes motion, then checks success and optionally records an episode. Supporting branches generate task programs and natural-language instructions, render or describe assets, and encode camera data for datasets. The excerpts show core runtime wiring, while policy-server internals and some persistence details remain unsampled.
Main entry points: _base_task.py Task Runtime · eval_policy_multitask.py Policy Evaluator · task_info.py Task Metadata · robot.py Bimanual Robot
Components 21
Task Environments
- Task Runtime
envs/_base_task.py - Task Scenarios
envs/place_bread_basket.py - Task Metadata
code_gen/task_info.py
Simulation Control
- Scene Actors
envs/utils/create_actor.py - Pose Utilities
envs/utils/transforms.py - Action API
envs/utils/action.py - Bimanual Robot
envs/robot/robot.py - Motion Planner
envs/robot/planner.py
Observation Data
- Camera Sensors
envs/camera/camera.py - Image Codec
data/decode_image_bit.py - Episode Files
envs/utils/save_file.py
Task Generation
- GPT Agent
code_gen/gpt_agent.py - Observation Agent
code_gen/observation_agent.py - Code Runner
code_gen/run_code.py - Instruction Generator
description/utils/generate_episode_instructions.py - Asset Renderer
description/utils/get_image_from_glb.py
Collection Evaluation
- Data Collector
scripts/collect_data.py - Policy Evaluator
scripts/eval_policy_multitask.py
Other components
- User
- Robot Policy
- XPolicyLab
Connections 22
- User → Task Runtime: selects task
- User → Policy Evaluator: starts evaluation
- Task Metadata → GPT Agent: guides generation
- GPT Agent → Code Runner: generates code
- Observation Agent → GPT Agent: returns observations
- Code Runner → Task Runtime: runs task
- Task Runtime → Scene Actors: creates actors
- Task Runtime → Camera Sensors: captures observations
- Task Runtime → Bimanual Robot: executes controls
- Task Scenarios → Action API: requests actions
- Action API → Bimanual Robot: dispatches actions
- Bimanual Robot → Motion Planner: plans motion
- Camera Sensors → Image Codec: encodes images
- Image Codec → Episode Files: writes frames
- Task Runtime → Episode Files: records episodes
- Data Collector → Task Runtime: collects episodes
- Policy Evaluator → Task Runtime: runs policies
- Policy Evaluator → Robot Policy: requests actions
- Robot Policy → Task Runtime: returns actions
- Policy Evaluator → XPolicyLab: uses deployment
- Instruction Generator → Task Metadata: renders instructions
- Asset Renderer → Task Metadata: visualizes assets