Flash ALOHA hits human-speed robotic manipulation

Most imitation learning setups leave robots moving like sluggish animations because teleoperation interfaces cannot capture high-speed demonstrations accurately. Flash ALOHA solves this by combining low-friction physical teaching with algorithmic trajectory smoothing, closing the speed gap between human intent and robotic execution.

• –Kinesthetic demonstration pipeline: Replaces clumsy puppeteering rigs and drift-heavy visuo-inertial tracking with backdrivable arm guidance, capturing natural fast-motion trajectories directly from human hands.
• –Action chunking stabilization: Incorporates lightweight trajectory smoothing to eliminate high-frequency tracking instabilities common in chunked ACT and Flow policy execution.
• –Operator visual masking: Dynamically strips demonstrator hands and bodies from training frames, allowing vision-only RGB policies to generalize without visual human interference.
• –High sample efficiency: Achieves >80% success rates across precision tasks like ball catching, shorts folding, and table bussing with only 50 demonstrations per skill.
• –Broad policy and hardware support: Runs on accessible ARX arms (L5 and Trossen_L5) with out-of-the-box compatibility for LeRobot Flow and OpenPI models.