OSU open-sources FAME-QP, a GPU QP solver cutting MPC latency for humanoid and quadrotor control
ohiostate · hf · 2026-11-12
Ohio State released FAME-QP, an open-source GPU-accelerated solver for dense quadratic programs in linear MPC. It pairs constraint-dependent operator representations with fused single/two-kernel iterations replayed via CUDA graphs, auto-selecting execution config through device calibration, probing a cached operator bank for initial penalties, and updating operators lazily. Benchmarks plus simulated Unitree G1 humanoid recovery (290-variable QP solved every 2ms after an 80N push) and Skydio X2 quadrotor tracking show lower latency than competing GPU solvers with unchanged accuracy. arXiv:2611.13766.
More from Embodied
- CoRL spotlight: test-time gradient guidance tackles RL scaling woes for flow policies — chris_j_paxton · 2026-11-12
- GXO Logistics partners with Apptronik to scale Apollo humanoid robots in operations — KyleMorgenstein · 2026-11-12
- CoRL 2026 R2RL workshop crowdsources a roadmap for sample-efficient real-robot RL — abhishekunique7 · 2026-11-12
- CMU Robotics Institute is hiring faculty across all tracks this year — davheld · 2026-11-12
- UR5 robot autonomously threads a cable through a car door with ACT policy — stepjamUK · 2026-11-12
- Apple's AMP1 hits #1 on NVIDIA RoboLab leaderboard by writing Python to control robots — chris_j_paxton · 2026-11-12