GATO documentation#
GATO is a GPU batched trajectory optimizer and MPC controller for robot arms and floating-base robots: a batched SQP with PCG or block-Cholesky linear solves, rigid-body dynamics generated per robot by GRiD, numpy, torch, JAX and C++/CUDA interfaces, and a constraint layer that keeps plans inside the URDF limits. The project page presents the paper and the latest results; this site is the reference: how to install and use the library, what it guarantees, and what it does not.
pip install gato-mpc, the first solve, the controller, the constraint layer,
adding a robot, the C++/CUDA interface.
Fourteen numbered scripts from one solve to MuJoCo in the loop; the paper figures and the benchmark harnesses.
The conventions an integration depends on: frames, limits, cost semantics, the solve interface, array kinds, API stability.
What is attested, what is experimental, the operating limits, and what the receipt proves.
Mechanisms, measured defaults and provenance for every row-group method: boxes, cones, collision, contact rows, the ADMM inner loop.
The current GPU test receipt: commit, outcome, module set, signer, and how to verify it yourself.
BSQP, MPCController, MPCPolicy, the hypothesis batch, tasks, worlds,
the builder and the fingerprint, from the docstrings.
Build, tests, the receipt procedure, timing windows, releasing.