repo · node map · vlalab 0.1
robot-learning: the VLA pipeline as swappable nodes
A working base for each of the eight stages in the VLA blog. Every stage is a node, every paper is a component you plug into a node, and everything is configured from YAML or Python. Pick a stage, a component or a pipeline node below, and the files that implement it light up in the tree.
$ git clone https://github.com/aPR0T0/robot-learning $ cd robot-learning && ./setup.sh # uv, CUDA torch, sims $ source .venv/bin/activate $ vlab run reach_smoke # 8 stages, ~3 min # swap a paper or a param from the CLI: $ vlab run reach_smoke vla.preset=vla/diffusion_tiny $ vlab run reach_smoke post.method=head_only
training stagedata stagedeploy / evallogged episodes feed back