Robot learning has a data problem. Expert demos are expensive, but the world is full of imperfect data: noisy teleop, sim demos, wrong tasks, cross-robot datasets. Ambient Diffusion Policy asks a better question: Not "should we use bad data?" but "when during diffusion is bad data still useful?"

The insight: Diffusion Policies learn actions coarse-to-fine. High noise → global task structure Low noise → local motion primitives So suboptimal data is only used where it helps: high noise when differences disappear, low noise when useful skills can still transfer. Everything else is filtered out.

The best part is how simple it is. No architecture changes. No inference changes. No new training loop. Just change the data sampler. On noisy demos, sim-to-real, task mismatch and Open X-Embodiment scaling, Ambient learns from imperfect data instead of being corrupted by it.

website:

The world produces far more imperfect data than perfect data and most of the time, the winners are the ones who figure out how to learn from it. Glad to see that we are moving in that direction too with robotics!