VLASH Training

Fine-tuning a state-of-the-art robot policy like π₀.₅ on your own task requires multi-GPU distributed training infrastructure that has never been publicly released for the VLASH framework. This project removes that barrier.

It packages VLASH — a VLA fine-tuning and deployment framework built on LeRobot — into a single container image that runs on a cloud VM, an HPC cluster, or a local workstation with one command. You bring your robot demonstrations; the pipeline handles the rest.

Workflow

1. Collect demonstrations on your robot  →  upload to HuggingFace Hub
2. Run:  ./scripts/train.sh config.yaml  →  fine-tuned checkpoint on HF Hub
3. Load checkpoint on your inference hardware  →  deploy

What VLASH adds over standard LeRobot

Feature	Description
Asynchronous inference	Temporal Delay Augmentation — 29.5× lower latency on embedded hardware
Memory-efficient fine-tuning	LoRA + shared observation encoding — fine-tune π₀.₅ on 12 GB VRAM
Distributed training	DeepSpeed ZeRO-2 (LoRA) and FSDP (full fine-tuning)
Portable containers	Same Docker/Singularity image on AWS, GCP, NSCC ASPIRE, or a local GPU

Validated end-to-end

The pipeline is validated on a ball pick-and-place task using a Piper arm with inference on a Jetson AGX Orin at 30 Hz:

• 65% task success rate with async inference vs. 5% synchronous baseline
• 29.5× latency reduction — 184.5 ms vs. 5444.1 ms per inference call
• 2.9× faster task completion — 52 s vs. 153 s

Quick start

# 1. Set required environment variables
export SCRATCH=/your/persistent/storage   # mounted disk
export HF_TOKEN=hf_xxx                    # HuggingFace token
export DATASET_REPO_ID=your-org/your-dataset

# 2. Pull the container
singularity pull vlash.sif docker://frieddeli/vlash-forge:latest
# or: docker pull frieddeli/vlash-forge:latest

# 3. Train
./scripts/train.sh examples/train/pi05/cloud.yaml

See First-time Setup for prerequisites.