Researchers led by Jianglong Ye have released RoboKit, a toolkit for running many differentiable optimisations in parallel for robotics, according to a paper accepted at CoRL 2026 and posted on arXiv. Its GPU solvers run several starting points at once and were 2 to 60 times faster than task-specific baselines across seven tasks, the authors report.
Coarse-to-fine representations speed up collision checks, rendering and simulation, so RoboKit optimised 4,096 grasps in 2.97 seconds. Because its costs are differentiable and its simulation parallel, it can also turn optimisation into learned policies, which ran 9 to 240 times faster at comparable accuracy, and train reinforcement learning, which reached a GPU baseline's scores up to 2.1 times sooner.



