Humanoid robots for last-mile delivery? Reward AI's OM-1 shows some cool tricks

Humanoids for last mile deliveries

Six weeks ago, Reward AI came out of stealth with a robot brain it says learned entirely from humans. Today it showed that brain doing delivery work.

In a new video posted this morning, a humanoid running OM-1 pulls packages out of the bed of a Tesla Cybertruck, carries them to a building entrance, and stacks them by the door. In a second clip, humanoids climb stairs to retrieve packages and carry them back down. 

Both are labeled "fully autonomous, 1x speed."

"We stress test OM-1 on last-mile delivery for packages from Amazon, Walmart, DHL, etc in different positions and orientations," the company posted on X. "Our model can reactively adjust, and let humanoids climb, navigate together with whole-body approaching, and hold packages firmly while climbing down."

(The Cybertruck, I suspect, is not an accident. If you want robot-delivery footage to travel on X, it doesn't hurt to put a Tesla in it.)

OM-1: data from humans

Quick refresher on what OM-1 is, because the approach is unusual.

Plenty of robot foundation models are trained on teleoperation: humans puppeting robots, thousands of hours at a time. Reward says OM-1 uses none of that. No teleop data, no on-robot data at all. 

Instead, people wear a sensorized seven-degree-of-freedom glove called the Omnibody Hand while they work, and it records hand pose, touch, proximity and vision. OM-1 learns from that, and a separate control layer trained with reinforcement learning in simulation translates its actions into joint commands for whatever body it's running on: a tabletop arm, an industrial arm, or a humanoid. Reward claims the model can pick up a new task from under 30 minutes of human data.

I flagged the launch in September as one of a cluster of companies, Figure and XPENG among them, chasing the same prize: getting human data to transfer to robots at scale.

The team has the résumé for it. CEO Zipeng Fu did his Stanford PhD with Chelsea Finn and worked on Mobile ALOHA and HumanPlus. CTO Chen Wang did his PhD with Fei-Fei Li and Karen Liu, and built DexCap, the portable hand-capture system the Omnibody Hand descends from.

Putting it all together

So what's new today?

The September launch was mostly arms on tables: bartending, folding laundry, packing phones, unplugging an Ethernet cable. The humanoid footage was mostly a robot opening a fridge to grab a drink. At the time, Fu said whole-body manipulation and navigation on humanoids would come later. This is that later: walking outdoors, carrying awkward loads, coordinating legs and arms on stairs, which is about the hardest place to hold a box.

"It was really exciting to see our humanoids pick up a package from the stairs and open the fridge to grab a drink," Yifeng Zhu, a member of Reward's technical staff, posted.

Last-mile delivery is a natural showcase. It's the part of logistics where wheeled robots and drones run out of road, because doorsteps come with stairs, curbs and porches. It's also where a lot of humanoid hopes are pinned. Agility's first customer just came back for more, but that's warehouse work. Doorsteps are a messier problem.

This is cool, but there's a few caveats still

Now the caveats, which are the same ones that hung over the launch.

Reward still hasn't published a paper, model weights, benchmarks or success rates. The only hard number in the September announcement measured the capture glove, not the policy: electromagnetic tracking cut hand-tracking overshoot from 24.9 mm to 9.5 mm at high speed. These are company-filmed demos, and a demo tells you a robot can do something, not how often it does it. The company hasn't disclosed funding, customers, or even whose humanoid this is; the robot in the video is dressed in work clothes and a hi-vis vest with a custom sensor head.

To be fair, Reward has been strict about one thing other demo reels often aren't: every clip is real-time. The company made a point of it at launch, telling viewers that if other models look faster, they should check the corner of the frame for a 2X or 5X label. That counts for something when the pitch is human-speed robots.

The bigger question is whether learning from humans, with no robot data at all, actually scales better than teleop, simulation or video prediction. Plenty of smart, well-funded people are betting on each. Machenta just raised $40 million on simulation, and Dream Labs raised $90 million on video prediction.

Reward's answer, for now, is a humanoid in a safety vest carrying a DeWalt box up to a door.

Sources

• Reward AI on X: OM-1 humanoid delivery demo (Oct. 27)
• Reward AI on X: OM-1 launch (Sept. 14)
• Reward AI: OM-1 technical blog post
• Forbes: Figure's billion-dollar physical AI bet