How huytai1o2/smolvla works
This repository packages a reproducible LIBERO orange-juice robot rollout: install the evaluation environment and fine-tuned policy, then run a closed-loop episode and inspect its success metrics and video. The documented workflow spans setup, policy/simulation interaction, and result generation. Relationships below are based on the README only; no script source excerpts were available, so implementation details beyond those documented are omitted.
Main entry points: install.sh Environment Installer · run_orange_juice.sh Episode Runner
Components 12
Local Setup
- Environment Installer
scripts/install.sh - LeRobot Runtime
- Fine-tuned Policy
- LIBERO Assets
Robot Rollout
- Episode Runner
scripts/run_orange_juice.sh - LIBERO Evaluator
- Robot Simulation
- Object Task 9
Evaluation Results
- Episode Metrics
- Raw Rollout Video
Other components
- User
- Hugging Face
Connections 15
- User → Environment Installer: runs setup
README.md - Environment Installer → LeRobot Runtime: installs
README.md - Environment Installer → Fine-tuned Policy: downloads
README.md - Environment Installer → LIBERO Assets: downloads
README.md - Hugging Face → Fine-tuned Policy: hosts checkpoint
README.md - User → Episode Runner: starts episode
README.md - Episode Runner → LIBERO Evaluator: runs evaluation
README.md - LIBERO Evaluator → Object Task 9: evaluates
README.md - LIBERO Evaluator → Robot Simulation: runs rollout
README.md - LIBERO Evaluator → Fine-tuned Policy: queries actions
README.md - Robot Simulation → Fine-tuned Policy: provides observations
README.md - LIBERO Evaluator → Episode Metrics: writes results
README.md - LIBERO Evaluator → Raw Rollout Video: generates video
README.md - User → Episode Metrics: inspects success
README.md - User → Raw Rollout Video: views rollout
README.md