Baseline VLA¶

Train and evaluate AlphaBrain's three base VLA frameworks — PaliGemmaOFT, PaliGemmaPi05, LlamaOFT — on LIBERO.

Prerequisites¶

conda activate alphabrain
cp .env.example .env                      # fill in the paths below

PRETRAINED_MODELS_DIR=data/pretrained_models
LIBERO_DATA_ROOT=/path/to/lerobot/libero
LIBERO_HOME=/path/to/LIBERO
LIBERO_PYTHON=/path/to/miniconda3/envs/libero/bin/python

See Installation for the full environment setup.

Train¶

# PaliGemmaPi05 multi-task (4 GPU, BS=256, 60k steps)
bash scripts/run_base_vla/train.sh paligemma_pi0_openpi_aligned_v3

# PaliGemmaOFT multi-task (4 GPU, BS=128, 150k steps)
bash scripts/run_base_vla/train.sh paligemma_oft_all_150k

# LlamaOFT multi-task (4 GPU, BS=128, 1.2M steps; LM frozen)
bash scripts/run_base_vla/train.sh llama_oft_all_150k

# Single-task examples
bash scripts/run_base_vla/train.sh paligemma_oft_goal
bash scripts/run_base_vla/train.sh llama_oft_goal

Checkpoints: results/training/<run_id>/checkpoints/steps_*.

Evaluate¶

bash scripts/run_base_vla/eval.sh paligemma_pi0_v2_goal_eval
bash scripts/run_base_vla/eval.sh paligemma_oft_bs128_goal_eval
bash scripts/run_base_vla/eval.sh llama_oft_eval

Results: per-task + overall success rate.

Next Steps¶

NeuroVLA — brain-inspired SNN action head
RL-Token — online RL fine-tuning
World Model — V-JEPA / Cosmos / Wan backbones
Continual Learning — sequential finetuning across 10 tasks

Full mode list, perf table, and custom-config overrides: scripts/run_base_vla/README.md.