Flash ALOHA: robots learn fast, human-speed bimanual skills from 50 demos
erwincoumans · x · 2026-11-14
Chelsea Finn, Zipeng Fu and colleagues released Flash ALOHA (first shown as ALOHA Lightning at ICRA), a system for fast and precise robot manipulation, with paper, code and hardware all public. It tackles slow teleop data and tracking failures at speed by using kinesthetic teaching to collect near-human-speed demos on a backdrivable bimanual platform.
The learning pipeline adds a lightweight smoothing algorithm to fix action-chunking jitter, and masks out the operator from camera views for visual alignment; the page also quantifies how visual-inertial tracking degrades on high-speed handheld trajectories.
With just 50 demos per task, Flash ALOHA completes bimanual tasks like ball tossing/catching, folding shorts, battery insertion and bussing tables at over 80% success rate, at close to human speed.
More from Embodied
- Xperience 2 turns real-world captures into interactive robot simulation environments — liuziwei7 · 2026-11-14
- OpenAI shows off limited-edition ModRetro Chromatic handheld in black collector's box — OpenAIDevs · 2026-11-14
- Tesla Optimus plus Starlink could put robots in remote sites before full autonomy arrives — jeasinema · 2026-11-14
- SpaceX Starship to attempt first full-orbit flight with payload deploy and tower catch — hey_abusiddik · 2026-11-14
- Simulated Body Recovers From Halved Leg Strength, Retrains Itself to Walk — MaziyarPanahi · 2026-11-14
- Starlink V3 satellite camera captures stunning view of Starship Ship 42 above Earth — XFreeze · 2026-11-14