Abstract
Learning from human demonstrations has enabled robots to acquire a wide range of manipulation skills, but learned policies typically execute far slower than ordinary humans. This speed gap is mainly due to lack of an interface for precisely collecting demonstration data at high speed, and the difficulty in training policies that can robustly execute high-speed motions. In this paper, we present Flash ALOHA, a practical system for learning fast and precise robotic manipulation. Our system uses kinesthetic teaching to intuitively collect accurate and near-human-speed demonstrations on a backdrivable bimanual platform, yielding natural and fast trajectories. Our learning pipeline consists of two components: a lightweight smoothing algorithm to address tracking instabilities produced by action chunking, and for visual alignment we mask out the operator visible in the camera views. Given 50 demonstrations for each task, Flash ALOHA autonomously completes bimanual tasks such as ball tossing and catching, folding shorts, battery insertion, and bussing tables at over 80% success rates while close to human speed.
Autonomous Skills
Speeding Up Puppeteering
Handheld Devices Tracking Error
Testing the pose estimation abilities of a state of the art visuo-inertial tracking device mounted on Flash ALOHA and replaying a human speed trajectory, we find that tracking fidelity degrades at high speed.
End effector (ground truth)Visual-inertial estimate→Error
Both videos and both graphs share one timeline. Scroll over a graph to zoom, drag to pan, click to jump to that moment, and double-click to reset.
Simple and practical setup
Drag to rotate
Right-hand scissor handles · drag to rotate