Transformers accelerate space robot trajectory planning by 23%

A team from Stanford University (Yuji Takubo, Maximilian Adang, Mac Schwager, Simone D'Amico) has published a paper on arXiv proposing a transformer-based warm-starting technique for sequential convex programming (SCP) in the terminal approach phase of a space manipulator toward a tumbling target. The problem of real-time trajectory generation for on-orbit robotic servicing is notoriously difficult due to the nonlinear coupling between spacecraft bus motion, manipulator dynamics, visibility cone, and safety constraints. The framework decomposes the problem into a system center-of-mass translational planning stage and a coupled attitude-manipulator torque-allocation stage, with the transformer warm-start applied to the latter (the computational bottleneck).

Over 300 held-out scenarios, the learned warm-start reduces the second-stage SCP iteration count by up to 28% and runtime by 23% while preserving the final control-cost distribution. For nonconvex feasibility projection, the warm-start nearly halves runtime relative to cost-optimal SCP and avoids catastrophic high-cost behavior seen with heuristic initializations. The authors compared linear and flow matching action decoders under different action-chunking and training dataset sizes. These results demonstrate that sequence-model warm-starts can significantly improve both computational efficiency and trajectory robustness for optimization-based terminal guidance in space manipulation, potentially enabling faster and safer autonomous satellite servicing.