NeuroVLA¶

Brain-inspired VLA: pretrain a Spiking Neural Network action head with backprop, fine-tune with hybrid R-STDP, evaluate on LIBERO.

Prerequisites¶

conda activate alphabrain
# Same .env fields as Baseline VLA

Hardware: 4 × A800 80 GB default. Smaller setups work with --gpus N.

Pipeline¶

1. Pretrain¶

bash scripts/run_brain_inspired_scripts/run_neurovla_pretrain.sh \
    --steps 50000 --run-id my_pretrain

2. R-STDP Fine-tune¶

bash scripts/run_brain_inspired_scripts/run_stdp_finetune.sh \
    --pretrained results/training/my_pretrain/checkpoints/steps_50000 \
    --steps 10000 --run-id my_stdp_ft

3. Evaluate¶

# libero_goal, 10 trials/task
bash scripts/run_brain_inspired_scripts/run_eval_libero.sh \
    --pretrained results/training/my_stdp_ft/checkpoints/steps_10000

# all 4 suites, 50 trials, with online STDP test-time adaptation
bash scripts/run_brain_inspired_scripts/run_eval_libero.sh \
    --pretrained results/training/my_stdp_ft/checkpoints/steps_10000 \
    --suite all --trials 50 --online-stdp

Results: results/evaluation/brain_inspired_eval_<timestamp>/<suite>/.

Full CLI reference and the team's open-research notes: scripts/run_brain_inspired_scripts/README.md.