On 27 October, Reward AI published new footage of its OM-1 humanoid — last-mile delivery, picking objects on stairs and household chores in a single video. More consequential than the scenes themselves is the claim about data sources: the company says its control policy uses neither teleoperation nor experience gathered on a robot.
Key takeaways
- The video covers a delivery, picking a parcel and a sock on stairs, mopping and opening a refrigerator door
- The company describes the footage as fully autonomous and played at normal speed
- The OM-1 policy trains on human demonstrations captured through the wearable Omnibody Hand
- A low-level controller learns by reinforcement learning in simulation, on a cycle separate from the policy
- Claimed threshold for a new task: under 30 minutes of human demonstration data
What does the video show?
The video opens with a last-mile delivery segment — the humanoid climbs onto the open tailgate of a Tesla Cybertruck to retrieve a parcel. From there, per the chapter list the company published, comes package picking on stairs at 00:28, mopping at 00:43, sock picking on stairs at 00:51 and refrigerator opening at 01:06.
The task selection is not arbitrary, and it is what makes the demonstration interesting. A package, a loose sock and a fridge door are three different manipulation problems: supporting a load, grasping deformable material and applying force against resistance.
Stairs add a fourth constraint, one that has nothing to do with the gripper. They change where the robot can stand to reach an object, because a step limits foot placement and shifts the centre of mass.
The delivery segment itself deserves attention. Climbing onto a tailgate is not a grasp but locomotion on an unstable, raised support, with manipulation following only afterwards. Most humanoid demonstrations keep those two apart — the robot either walks or picks something up while standing firmly on a floor.
Reward AI's site labels the footage fully autonomous and played at 1x speed. That is the manufacturer's description, not an independent evaluation of reliability.
Where does OM-1 get its data?
In its technical overview the company describes OM-1 as a shared policy trained on human demonstrations captured through the wearable Omnibody Hand. The device combines hand-motion tracking with visual, tactile, proximity and force information.
The aim is to preserve how people actually handle objects, including the contact cues ordinary video cannot capture. Footage shows where a hand was. It does not show how hard the fingers squeezed.
The company states the policy uses neither teleoperation nor on-robot experience as training data. That does not mean the whole system is built without simulation: a low-level controller learns through reinforcement learning in simulation and translates policy outputs into motions suited to a particular robot's dynamics.
That controller runs independently of the policy's inference cycle, with transitions between consecutive predictions smoothed to keep motion continuous. Reward AI also states a new task can be learned from less than 30 minutes of demonstration data. That is a claim about demonstration effort, not about total training time or the amount of data behind the general model.
This is a simplified path built on the manufacturer's technical description, not an independently verified architecture.
Why not just record the robot through teleoperation?
The most common way to gather manipulation data today is teleoperation: a human drives the robot while the system records observation-action pairs. The advantage is obvious, because the data arrives already in the format the robot will act in.
So is the cost. Every hour of data takes an hour of operator time on a specific robot, so the dataset grows linearly with the number of machines and stays tied to one design.
A glove demonstration inverts that dependency. A person performs the task with their own hands, so data collection needs no robot and is not capped by how many exist. In exchange comes a problem teleoperation does not have: a human hand has different kinematics and a different force range than a gripper.
And that gap is what the two-layer design is meant to bridge. The policy learns what in a task is body-independent — ordering, contact points, intent. The low-level controller, trained in simulation and carried across to hardware sim-to-real, handles the fit to the machine's actual mechanics. This belongs to the cross-embodiment family of approaches, meaning transfer of skills between different bodies.
What do these names mean?
Three terms needed to judge the claim:
- Policy — a model that picks the next motion from an observation. The equivalent of a function taking state and returning an action, only learned rather than hand-written.
- Omnibody Hand — a wearable glove recording hand motion together with vision, touch, proximity and force. It is a data-collection tool, not part of the robot.
- Low-level controller — the layer translating policy outputs into motor commands for a specific machine. The equivalent of a device driver between an application and hardware.
Separating the policy from the controller is the whole point here: task knowledge can be trained without a robot, while the fit to hardware is left to a separate layer that is cheaper to train.
Is it working yet?
This is a demonstration, not a product or a deployment. The 27 October video shows specific behaviours, but on its own it does not establish how often those tasks succeed, how many attempts were recorded, or how performance holds up with unfamiliar objects and locations.
For delivery and household work those questions weigh as much as a successful sequence. Repeated trials across different stairways, packages and door mechanisms would show how far the demonstrated skills extend beyond the selected clips.
The company emerged from stealth last month, presenting its approach to transferring human manipulation skills across different robot bodies. This video extends that announcement rather than marking a new stage.
The second thing this video shows in passing is a shift in scene selection. Delivery and household work are tasks where the environment has not been prepared for a robot: stairs vary in height, parcels in size, and fridge doors in resistance. That is a harder testing ground than a station on a shop floor, though without published success rates it is hard to say how much harder.
Why it matters
The bottleneck in manipulation today is not hardware but data. If the threshold for learning a new task really drops to half an hour of demonstration without a robot involved, the economics of the whole process change: the dataset stops scaling with the fleet. It is a manufacturer's claim, so it gains value only once someone reports success rates across repeated trials.
What's next
- The missing data point is success metrics — the number of attempts and the share of successful runs for each demonstrated task
- The second checkable step is generalisation: the same tasks on unfamiliar stairs, packages and door mechanisms
- The claimed under-30-minute threshold covers a new task, not the total training of the general model — only a description of the full dataset will settle that
Sources
- Humanoids Daily — Reward AI Shows OM-1 Handling Deliveries, Stairs and Household Chores
- Reward AI — OM-1 technical overview
- YouTube — Reward AI OM-1: delivery, stairway manipulation and household tasks





