Trajectory optimization for robots often requires many evaluations of functions derived from the robot's dynamical model. These evaluations support a range of methods, from sampling-based zero-order approaches to gradient-based first- and second-order techniques. The new arXiv preprint introduces CusADi2, a library designed to perform these computations in batches on the GPU, with a focus on minimizing latency.

CusADi2 specifically targets robot kinematics, Jacobians, and Hessians, which are core building blocks in many optimization pipelines. By offloading batch computation to the GPU, the library aims to reduce the time spent on repeated function evaluations, a common bottleneck in trajectory optimization. The paper's emphasis is on latency optimization rather than raw throughput, which could make it particularly useful for real-time or near-real-time control applications.

The source is a single arXiv preprint, so the claims are not yet peer-reviewed. The abstract does not provide benchmark results or implementation details, but the stated goal is clear: faster, batchable computation of robot dynamics on GPU hardware. Further reading of the full paper would be needed to assess actual performance gains and trade-offs.