Learned only from humans: Reward AI shows humanoid mopping and fetching parcels
A new video from the start-up shows its OM-1 robot model on deliveries, on stairs and doing chores. It was trained only on recordings of people wearing a sensor glove.
Photo: NatiSythen / Wikimedia Commons, CC BY-SA 3.0 (cropped and resized)
Video: YouTube · Reward AI · youtube.com
Robotics start-up Reward AI on Tuesday released a new video of its OM-1 control model. It shows a humanoid robot taking a parcel from the bed of a pickup truck, picking up packages and socks on a staircase, mopping the floor and opening a fridge. According to the company, the robot works fully autonomously and the footage runs at normal speed.
What sets it apart is the training: OM-1 learns solely from demonstrations by people wearing a wearable sensor glove called the Omnibody Hand. It records camera images, touch pressure, finger distance just before contact, the hand's path and forces. The team says it uses neither teleoperated robot sessions nor data from robots. The robot is said to learn a new task from less than 30 minutes of human demonstration, and the model is meant to run on industrial arms as well as humanoids.
Reward AI unveiled OM-1 in mid-September. The start-up is run by researchers who previously worked on the DexCap, HumanPlus and ALOHA projects. No success rates, model weights or independent evaluation have been published so far, and it is not known how many attempts the scenes shown required.