Amortizes iterative diffusion into a one-step trajectory policy for robotics using a novel 'Keyed Drift Field' objective.
March 17, 2026
Original Paper
Amortizing Trajectory Diffusion with Keyed Drift Fields
arXiv · 2603.14056
The Takeaway
Iterative diffusion denoising is often too slow for high-frequency closed-loop control in robotics. This method enables one-step inference that retains the multi-modality and diversity of diffusion, allowing high-performance robotic planning on limited compute budgets.
From the abstract
Diffusion-based trajectory planners can synthesize rich, multimodal action sequences for offline reinforcement learning, but their iterative denoising incurs substantial inference-time cost, making closed-loop planning slow under tight compute budgets. We study the problem of achieving diffusion-like trajectory planning behavior with one-step inference, while retaining the ability to sample diverse candidate plans and condition on the current state in a receding-horizon control loop. Our key obs