GaP — Graph-as-Policy#
🧪 GaP Beta Code release (1 July 2026)
GaP is under active development and now in beta testing. Please send comments and suggestions to kych@berkeley.edu — we plan to release an updated version by 1 Aug 2026. Expect rough edges: APIs, the workflow schema, and skill interfaces may change without notice between releases.
The graph is the policy. GaP is a multi-agent coding harness that compiles a natural-language task into a typed, verified computation graph of modular skills, self-improves it in simulation, and runs the graph — not a black-box policy — on simulators and real robots. It targets Variational Automation (VA): tasks a robot must perform persistently and reliably across many varying instances (objects vary in geometry and pose), not just solve once.
GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness for Variational Automation
Real-robot rollouts of graphs generated from one-sentence task descriptions.
The Overview has the full story — what Variational Automation is, why the policy is a graph, and what you need to run GaP.
Hello world — your first run#
The fastest path is the 15-minute tour: clone the two repos side by side, sync, install the skill bundles, set one LLM key, and run the hero LIBERO graph end-to-end with the trace open.
git clone --recurse-submodules https://github.com/graph-robots/graph-as-policy.git
git clone https://github.com/graph-robots/open-robot-skills.git # sibling, auto-discovered
cd graph-as-policy
uv sync # engine + LIBERO sim
uv run gap skills install --all # per-bundle venvs (sam3, cuRobo, vlm, …)
export HF_TOKEN=... # gated SAM3 weights
export OPENROUTER_API_KEY=... # codegen + in-graph VLM (openrouter is the default)
MUJOCO_GL=egl uv run gap run examples/libero_quickstart/graph --sim libero_object_all_variance/0
uv run gap viz # browse the trial at localhost:9432
uv sync + gap skills install --all — engine, sim, perception, motion planning, each in its own venv.
Zero to a verified rollout on LIBERO — then a packing loop with a real backward edge.
Eleven examples, from the end-to-end quickstart to the release gate and real robots.
Contributable perception, grasping, transport, tracking, and policy skills.
gap run, generate, check, registry, skills, benchmark, viz, and more.
One plugin install teaches AI coding agents to drive the whole workflow.
Read more in Overview, or jump straight to Installation.