Differentiable policy trajectory optimization with generalizability (DiffOG). Visuomotor policies enhanced by DiffOG generate smoother constraint-compliant action trajectories in a more interpretable way. DiffOG introduces a novel transformer-based differentiable trajectory optimization framework tailored for action refinement in imitation learning. Leveraging the differentiability of the optimization layer and the high capacity of the transformer, DiffOG can be trained on demonstration data to adapt to the diverse characteristics of trajectories across different tasks. We evaluate DiffOG across 13 tasks and showcase four representative ones here. These selected tasks present several key challenges, including long-horizon dual-arm manipulation, high-precision control, and smooth constraint-satisfying trajectory generation.
DiffOG, a differentiable #TrajectoryOptimization layer that enhances visuomotor policies by generating smoother, constraint-compliant action trajectories with better generalization & interpretability — improving performance over baseline methods.
https://ieeexplore.ieee.org/document/11267071
