
Improving Computational Efficiency for Powered Descent Guidance via Transformer-based Tight Constraint Prediction
Richard Linares of MIT presents this Stanford AA289 Robotics and Autonomous Systems Seminar talk on speeding up trajectory optimization for spacecraft landing. The problem: powered descent guidance solvers are too slow for flight-grade computers on autonomous spacecraft and surface robots. Linares introduces T-PDG, a transformer neural network trained on prior trajectory optimization runs that predicts the tight constraints defining the optimal landing trajectory and final landing time, using only spacecraft state and landing site parameters as input. Applied to the Mars powered descent guidance problem, the method cuts computation time for the fuel-optimal three degree of freedom trajectory by up to an order of magnitude compared to lossless convexification, while a built-in feasibility check preserves safety guarantees. The talk covers the attention mechanism's role in handling long time series sequences and positions the approach within the broader push for capable autonomy stacks on resource-constrained exploration missions.