How dstate/lbp works
LBP learns a LIBERO robot policy that plans backward from a latent final goal, predicts intermediate subgoals, then generates chunked actions conditioned on those goals. The README’s workflow is dataset conversion, planner training, policy training, and LIBERO evaluation; the graph focuses on runtime components and the policy-training entry point. Some planner-training and evaluation wiring is documented but its implementation was not sampled, so those links are limited to supported references.
Main entry points: train_policy_sim.py Policy trainer · LBP.py LBP policy · eval_libero.py LIBERO evaluator · LiberoEngine.py Dataset engine
Components 18
Data preparation
- HDF5 converter
hdf52hdf5.py - LIBERO trajectories
- Dataset engine
datasets/LiberoEngine.py
Latent planning
- Latent planner
models/MidPlanner.py - DecisionNCE encoder
Policy model
- LBP policy
models/LBP.py - Goal attention
models/components/CrossAttn.py - FiLM vision encoder
models/components/ResNet.py - Action heads
models/components/ActionHead.py - Residual MLP
models/components/MlpResNet.py
Training
- Policy trainer
train_policy_sim.py - Model wrapper
utils/wrapper.py - LR scheduler
utils/scheduler.py - Training logger
utils/logger.py - Checkpoint utilities
utils/tools.py
Robot evaluation
- LIBERO evaluator
eval_libero.py - LIBERO simulator
Other components
- Researcher
Connections 20
- Researcher → HDF5 converter: converts dataset
README.md - HDF5 converter → LIBERO trajectories: writes trajectories
README.md - LIBERO trajectories → Dataset engine: loads trajectories
- Researcher → Policy trainer: trains policy
README.md - Policy trainer → Dataset engine: creates data engine
train_policy_sim.py - Policy trainer → LBP policy: creates model
train_policy_sim.py - LBP policy → Latent planner: generates subgoals
models/LBP.py - Latent planner → DecisionNCE encoder: encodes image and language
models/MidPlanner.py - LBP policy → Goal attention: fuses goal latents
models/LBP.py - LBP policy → FiLM vision encoder: encodes observations
models/LBP.py - LBP policy → Action heads: predicts actions
models/LBP.py - Action heads → Residual MLP: uses residual network
models/components/ActionHead.py - Policy trainer → Model wrapper: wraps model
train_policy_sim.py - Policy trainer → LR scheduler: updates learning rate
train_policy_sim.py - Policy trainer → Training logger: logs metrics
train_policy_sim.py - Policy trainer → Checkpoint utilities: saves checkpoints
train_policy_sim.py - Researcher → LIBERO evaluator: evaluates policy
README.md - LIBERO evaluator → LIBERO simulator: runs benchmark
README.md - LIBERO evaluator → Dataset engine: imports dataset package
eval_libero.py - LIBERO evaluator → LBP policy: imports model package
eval_libero.py