Most modern visual imitation learning systems rely on multiple cameras, including wrist-mounted views, to capture fine interaction details. The authors of this paper argue that single-camera operation remains challenging because the policy must infer those details from one global view. They introduce Spatially Conditioned Diffusion Policy (SCDP), which treats end-effector trajectories as visual attention anchors that highlight task-relevant regions during the diffusion process.

SCDP combines a multi-scale visual encoder with a spatial conditioning module. The encoder produces feature maps at multiple resolutions, preserving both broad context and fine-grained visual cues. The conditioning module then samples point-wise features along intermediate end-effector trajectories within the diffusion loop, guiding the policy toward precise manipulation without needing wrist cameras.

In simulation, SCDP consistently outperformed strong single-view baselines and achieved performance comparable to multi-camera baselines. Real-world experiments demonstrated precise manipulation and robustness to visual distractors, suggesting that single-camera imitation learning can be a viable alternative to multi-camera setups. The paper was accepted at CoRL 2026.