Principle:ARISE Initiative Robomimic Environment Reconstruction
| Knowledge Sources | |
|---|---|
| Domains | Robotics, Evaluation, Simulation |
| Last Updated | 2026-02-15 08:00 GMT |
Overview
A metadata-driven environment reconstruction pattern that recreates simulation environments from checkpoint metadata to ensure evaluation conditions match training conditions.
Description
Environment Reconstruction creates simulation environments that match the exact conditions under which a policy was trained. This is critical for valid evaluation: the environment type (robosuite, gym, iGibson), task configuration, robot type, and rendering settings must all match.
The reconstruction reads env_metadata from the checkpoint, which contains the environment name, type identifier, and construction kwargs. It also inspects shape_metadata to determine if image or depth observations were used during training, automatically enabling off-screen rendering if needed.
The principle supports:
- Environment name override: Allow evaluation on different task variants
- Rendering mode selection: On-screen vs. off-screen rendering
- Environment wrapping: Apply wrappers (e.g., FrameStackWrapper) from config
Usage
Use this principle after checkpoint loading and before rollout execution. The reconstructed environment provides the simulation context for closed-loop policy evaluation.
Theoretical Basis
# Abstract environment reconstruction (not real implementation)
env_meta = checkpoint["env_metadata"]
shape_meta = checkpoint["shape_metadata"]
# Create env matching training conditions
env = create_env_from_metadata(
env_meta=env_meta, # Contains env name, type, kwargs
render_offscreen=shape_meta["use_images"], # Auto-enable for image policies
)
# Apply wrappers from config
env = wrap_env_from_config(env, config) # E.g., FrameStackWrapper