How nvlabs/vt-refine works
VT-Refine releases workflows for tactile-policy pre-training, evaluation, and simulation fine-tuning for bimanual assembly. The README identifies dppo/script/run.py as the entry point and documents the stages and data-fetch command. The tree names the easysim and easysim-envs submodules, but no source excerpts show their runtime interfaces, so their links are limited to the README’s documented simulation fine-tuning relationship.
Main entry points: dppo Fine-tuning · easysim Simulation runtime · fetch_vt-refine_data.sh Data fetch · easysim-envs Simulation tasks
Components 9
Policy workflows
- Data fetch
data/fetch_vt-refine_data.sh - Workflow runner
- Pre-training
dppo - Pretrain evaluation
dppo - Fine-tuning
dppo - Fine-tune evaluation
dppo
Simulation
- Simulation runtime
easysim - Simulation tasks
easysim-envs
Other components
- Researcher
Connections 8
- Researcher → Data fetch: fetches data
README.md - Researcher → Workflow runner: runs workflows
README.md - Workflow runner → Pre-training: runs pre-training
README.md - Workflow runner → Pretrain evaluation: runs evaluation
README.md - Workflow runner → Fine-tuning: runs fine-tuning
README.md - Workflow runner → Fine-tune evaluation: runs evaluation
README.md - Fine-tuning → Simulation runtime: uses simulation
README.md - Simulation runtime → Simulation tasks: uses tasks