Skip to content

Quickstart

From a blank machine to a trained-and-evaluated VLA checkpoint — then into any of the five capability-specific paths.

Read top to bottom

Work through Installation → Baseline VLA first. Everything below that is optional and can be tackled in any order.

The path

Step Page What it does Default hardware
0 Installation Conda envs, Flash Attention, pretrained weights, LIBERO data, .env config. —
1 Baseline VLA PaliGemmaOFT / Pi05 / LlamaOFT finetune + LIBERO eval — the first trained checkpoint from the default recipe. 4× A800
2 NeuroVLA Pretrain a brain-inspired VLA, then R-STDP fine-tune the SNN action head. 4× A800
3 RL-Token Off-policy TD3 RL on a pretrained QwenOFT (encoder pretrain + rollout + TD updates). 6× A800
4 World Model V-JEPA / Cosmos / Wan backbones × GR00T / OFT / PI decoders (12 combinations) + CosmosPolicy. 4× A800
5 Continual Learning Sequential finetuning across 10 LIBERO tasks with Experience Replay (LoRA + full-param). 2× A800

Every capability page follows the same layout — Overview → Prerequisites → Quick Start → Tips → Pointers — so you can skim one and recognize the shape of the next.

Minimum requirements

  • Python 3.10+
  • CUDA 11.8 / 12.x with matching PyTorch 2.1+
  • At least one A100/H100-class GPU (4× A800 recommended for the default Baseline VLA recipe)
  • ~100 GB of local disk for LIBERO datasets + pretrained VLMs

Detailed environment setup lives in Installation.

Shared conventions

Every quickstart assumes the same project layout:

AlphaBrain/
├── .env                               # local paths (see Installation)
├── configs/
│   ├── finetune_config.yaml          # single-entry training/eval modes
│   ├── continual_learning/*.yaml     # CL-specific configs
│   └── rl_recipes/*.yaml             # RL-specific configs
├── scripts/
│   ├── run_finetune.sh               # training launcher
│   ├── run_eval.sh                   # eval launcher
│   ├── run_base_vla/                 # Baseline VLA wrappers
│   ├── run_brain_inspired_scripts/   # NeuroVLA wrappers
│   ├── run_rl_scripts/               # RL-Token wrappers
│   ├── run_continual_learning_scripts/
│   └── run_world_model/              # World Model wrappers
└── results/
    ├── training/<run_id>/            # finetune outputs
    └── evaluation/<run_id>/          # eval outputs

All launchers:

  1. Source .env at the project root
  2. Read a mode block from YAML (mode name → resolved config)
  3. Create the output dir and snapshot the config for reproducibility
  4. Call accelerate launch (training) or start a server + client (eval)

Shared prerequisites

Run these once before any capability page:

  1. Complete Installation — main conda env + flash-attn + eval env + LIBERO data.
  2. Fill out .env with PRETRAINED_MODELS_DIR, LEROBOT_LIBERO_DATA_DIR, LIBERO_DATA_ROOT, LIBERO_HOME, LIBERO_PYTHON.
  3. Download the backbone(s) you need (see each page's Prerequisites section).

Which page should I read?

  • "I just installed — does training work?" → Baseline VLA
  • "How do I finetune a different backbone?" → Baseline VLA § Switch Backbone
  • "Is there a bio-plausible fine-tuning path?" → NeuroVLA
  • "Can I improve a pretrained VLA with RL?" → RL-Token
  • "I want the SOTA visual backbone." → World Model (Cosmos 2.0 + GR00T currently leads)
  • "How do I handle task drift across 10 tasks?" → Continual Learning

Getting help