HOME / PROJECTS / WIZARD
Embodied AI · Manipulation

WIZARD

A research note on one-shot imitation learning: how a manipulation policy can be taught from a single human demonstration and still generalise to tasks it has never seen, with no retraining step.

PROJECT PAGE ↗
~14×
success rate over the one-shot imitation baseline on our unseen-task evaluation set. Benchmark and protocol are in the paper.
1
video demonstration, then autonomy — no per-task teaching pass.
Collaboration
Developed in active collaboration with UC Berkeley. Co-authored, with a public project page.
Stage
Active research project, not yet in production. Validated on real hardware in the lab.
Why it matters
Teaching cost, not peak capability, is what keeps robots out of variable environments.
Wine bottle → cabinet
Cup pickup · real world

How it works

WIZARD watches a single human demonstration and extracts the underlying task structure — the sub-goals, contacts, and motions that make a task succeed — rather than memorising a trajectory. That representation is what lets it adapt instantly to new objects, positions, and environments.

Because the policy reasons about intent instead of replaying motion, it recovers from disturbances and generalises to tasks it has never explicitly been trained on — the behaviour behind the reported improvement on unseen tasks. What the figure does not yet cover is sustained operation in an industrial setting — that is the next stage of the work.