ALOHA Lightning: Learning Fast and Precise Manipulation
TL;DR Operators record near-human-speed demonstrations by physically backdriving a bimanual robot, and a learning pipeline masks out their hands and smooths the joins between action chunks. With 50 demonstrations per task, the policy buses a table in 6.8 s at 100% success, against 16.1 s and 90% for a policy trained on puppeteering data.
What they do
Learned policies usually move much slower than people, partly because teleoperation makes operators slow down. ALOHA Lightning uses kinesthetic teaching instead: the operator holds the grippers and moves the arms directly, near natural speed. The learning pipeline then removes the operator’s hands from the images and smooths the policy’s output.
How it works
- Hardware. Bimanual 6-DoF arms (ARX, and HESSIAN for the fast Toss & Catch) with planetary actuators that are easy to backdrive, gravity-compensated by torque control at 100 Hz. Printed thumb and finger handles on each gripper jaw; two wrist cameras and one top-down camera at 50 Hz.
- Masking. Forward kinematics and camera calibration locate the operator’s hands; regions around the handles are blacked out in every camera, in training data and live at deployment.
- Policy. ACT predicts chunks of future joint positions for a PD controller, from 50 demonstrations per task.
- Test-time smoothing. Consecutive action chunks overlap by a few steps and are blended with a sigmoid weight, which removes the jumps at chunk boundaries that grow with speed.
Results
- Against ACT on puppeteering data: Unstack Cups 90% in 3.97 s vs 75% in 8.84 s; Battery Insertion 85% in 5.70 s vs 55% in 17.87 s; Bus Table 100% in 6.81 s vs 90% in 16.05 s. Speeding up puppeteered policies instead drops Battery Insertion to 15-30%.
- Masking vs raw images: 90 vs 80% (cups), 85 vs 75% (battery), 100 vs 90% (shorts). Diffusion inpainting does worse than raw images on battery insertion (65%).
- Without smoothing, cup unstacking falls to 65%; temporal ensembling roughly doubles execution time.
- Toss & Catch takes 1.85 s; every toss succeeds and half the catches do. It could not be collected with puppeteering. In a seven-person user study, novices inserted the battery up to three times faster than with puppeteering.
Limitations
The mask is conservative and erases useful visual cues near the hands. The policy uses only vision and proprioception, with no touch, which rules out contact-rich dexterous tasks.
Videos
Figures




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 collecting demonstration data at high speed, and the difficulty in training policies that can robustly execute high-speed motions. In this paper, we present ALOHA Lightning (Flash ALOHA), a system for learning fast and precise robotic manipulation. Our system uses kinesthetic teaching to intuitively collect near-human-speed demonstrations on a backdrivable bimanual platform, yielding natural and fast trajectories. We also present a learning pipeline that enables smooth high-speed execution through test-time action smoothing and aligns the visual data distribution between data collection and deployment with masking. Given 50 demonstrations for each task, ALOHA Lightning autonomously completes bimanual tasks such as ball tossing and catching, folding shorts, battery insertion, and bussing tables for over 80% success rates while close to human speed. Project website: https://flash-aloha.github.io/
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
