How internrobotics/internw0-delta works
InternW0-Δ is a robot world–action model trained on robot and human demonstrations, then used to generate actions for benchmark simulators or real robots. Its primary workflow runs from dataset preparation through pretraining and task-specific post-training to online policy inference and evaluation or deployment. The model couples video and action experts with visual memory and language conditioning; 4D teacher supervision is training-only. The README documents the broad product stages; sampled source and the SOURCE INDEX clarify selected internal relationships, while unsampled implementation details are not inferred.
Main entry points: online_action_policy.py Online action policy · trainer.py Training loop · world_action_model.py World-action model · datasets Dataset loading
Components 20
Data preparation
- Dataset loading
src/wam/datasets - Video dataset
src/wam/datasets/lerobot/robot_video_dataset.py - Pretraining loader
src/wam/datasets/pretrain_lerobot_loader.py - Artifact cache
src/wam/cache/manager.py
Model and training
- World-action model
src/wam/model/world_action_model.py - Video expert
src/wam/model/backbones/wan22/wan_video_dit.py - Action expert
src/wam/model/modules/experts/action_dit.py - Language understanding
src/wam/model/modules/understanding/qwen_vl_encoder.py - 4D distillation
tools/4D_distillation/teacher.py - Training loop
src/wam/trainer.py - Model checkpoint
src/wam/training/checkpoint.py
Action inference
- Online action policy
src/wam/inference/online_action_policy.py
Evaluation
- LIBERO evaluation
eval/libero/eval_libero_single.py - LIBERO Plus
eval/libero_plus/server.py - RoboTwin evaluation
eval/robotwin/policy.py - RoboDojo evaluation
eval/robodojo/server.py
Robot deployment
- Deployment policy
deploy/policy.py - Robot controller
deploy/controller.py
Other components
- Robot demonstrations
- Real robot
Connections 21
- Robot demonstrations → Dataset loading: prepare data
README.md - Dataset loading → Pretraining loader: load sources
README.md - Dataset loading → Video dataset: prepare samples
src/wam/datasets/lerobot/robot_video_dataset.py - Pretraining loader → Training loop: train from data
src/wam/trainer.py - Training loop → World-action model: optimize model
src/wam/trainer.py - World-action model → Video expert: includes expert
README.md - World-action model → Action expert: includes expert
README.md - Video expert → Action expert: exchange features
README.md - World-action model → Language understanding: uses conditioning
README.md - 4D distillation → Training loop: supervise training
README.md - Training loop → Model checkpoint: save weights
src/wam/trainer.py - Model checkpoint → Online action policy: load weights
- Online action policy → LIBERO evaluation: predict actions
eval/libero/eval_libero_single.py - Online action policy → RoboTwin evaluation: predict actions
eval/robotwin/policy.py - Online action policy → LIBERO Plus: serve policy
eval/libero_plus/server.py - Online action policy → RoboDojo evaluation: serve policy
- Real robot → Deployment policy: provide observations
- Deployment policy → Online action policy: invoke policy
deploy/policy.py - Deployment policy → Robot controller: send actions
- Robot controller → Real robot: control robot
- Artifact cache → Training loop: cache artifacts
src/wam/trainer.py