Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation

Tony Tao, Jason Jingzhou Liu, Ruslan Salakhutdinov, Deepak Pathak

November 2026

TL;DR Pixel is a wheeled humanoid that costs under $15,000. The robot, its teleoperation interfaces, a wearable human-demonstration rig, the whole-body controller and the policy were designed together, so human and robot data share one action space. Policies trained only on wearable-rig demonstrations reach 82.0% average task progress on three household tasks, and collecting those demonstrations is 4.2-6.2x faster than teleoperation.

What they do

Pixel is a mobile-manipulation stack built around one low-cost wheeled humanoid. The robot, its data-collection interfaces, the whole-body controller and a hierarchical policy were designed together. As a result, teleoperated demonstrations and robot-free human demonstrations map to the same actions on the same robot. The hardware costs under $15,000 and is open source.

How it works

Robot. A 23-DoF platform: holonomic swerve base, 430 mm vertical lift, torso pitch joint, two 7-DoF arms with human-like kinematics and a 2-DoF head. The single-DoF pincer grippers have compliant TPU and silicone fingers. Cameras sit on the head and both wrists.
Data collection. Three teleoperation interfaces: a 65%-scale leader arm, VR with a joystick for the base, and VR with body tracking for the base. There is also a wearable rig under $1,000, with 386 g grippers that copy the robot’s fingers and wrist cameras and a VR headset that tracks the head. All of them produce head and gripper pose targets for one shared whole-body controller based on geometric fabrics.
Policy. A language-conditioned flow-matching transformer reads all three cameras. It predicts head and gripper poses in a “relative-base” frame placed on the ground under the head, which works for both human and robot data. A Gemini-based agent splits spoken requests into skills from a fixed library and monitors their progress.

Results

Average task progress is 82.0% from human demonstrations (toy cleanup, tidy pillow, trash disposal) and 78.7% from teleoperated ones (vacuuming, hanging a jacket, fetching a snack).
Relative-base actions reach 82.0%, vs. 67.3% for UMI-style relative actions and 25.7% for a head-anchored frame. With all joints active the system scores 100% on the three human-data tasks; disabling the torso, lift or head drops it to 47.4%, 71.2% and 52.0%.
In a 12-person study, every participant ranked the wearable rig first, and it was 4.2-6.2x faster than teleoperation. Adding teleoperated data to human data raised progress from 85.0% to 94.3%.

Limitations

Pixel cannot climb stairs and has no dexterous hands. The wearable rig records only head and gripper poses: no other body contact (such as bracing a door with an elbow) and no touch. The high-level agent was not systematically evaluated and cannot go beyond its skill library. On teleoperated tasks, joint-space actions scored higher (86.3%) than the chosen relative-base actions.

Figures

From Figure 2: Hardware co-design. The robot and the sensorized human data-collection system share camera viewpoints,... — From Figure 2: Hardware co-design. The robot and the sensorized human data-collection system share camera viewpoints, aligned human-robot kinematics, paired end effectors and a functionally aligned lower body.

From Figure 5: Policy overview. A high-level policy turns voice commands, prior context and keyframes into the next s... — From Figure 5: Policy overview. A high-level policy turns voice commands, prior context and keyframes into the next subtask. A low-level policy conditions on that subtask and the camera images and generates actions with a DiT-based action expert.

From Figure 6: Evaluation tasks. Toy cleanup, tidy pillow and trash disposal are trained from sensorized human data; ... — From Figure 6: Evaluation tasks. Toy cleanup, tidy pillow and trash disposal are trained from sensorized human data; vacuum cleaning, hang jacket and snack retrieval from robot teleoperation data. Tidy pillow and wine pouring are used in the user study.

From Figure 12: Observation- and action-space ablations for policies trained on (a) sensorized human data and (b) rob... — From Figure 12: Observation- and action-space ablations for policies trained on (a) sensorized human data and (b) robot teleoperation data.

Abstract

General-purpose mobile manipulation requires a combination of capable robot hardware, scalable demonstration collection, expressive whole-body control, and effective policy learning. Yet these components are often developed independently, creating mismatches that limit the capabilities of the complete system. We present Pixel, a co-designed full-stack system for learning general-purpose mobile manipulation. Pixel combines a low-cost wheeled humanoid with human-aligned kinematics and compliant grippers, multiple teleoperation and robot-free human data-collection interfaces, a unified whole-body controller, and a common policy learning pipeline across human and robot demonstrations. A hierarchical learning stack further enables language-conditioned visuomotor control and long-horizon task execution. We evaluate the system across hardware, data collection, and policy learning, and demonstrate mobile-manipulation behaviors requiring dexterity, active perception, navigation, and whole-body coordination. We hope that Pixel provides an integrated and extensible foundation for studying how robot hardware, data, control, and learning can be designed together for robot learning. More results at https://pixel-robot.com/pixel

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.

Matt Strong