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

TL;DR

Pixel presents a co-designed full-stack robot system integrating low-cost hardware, human-aligned kinematics, compliant grippers, and unified data collection and learning pipelines for general-purpose mobile manipulation. It enables hierarchical, language-conditioned visuomotor control and long-horizon task execution through a unified policy learning framework across human and robot demonstrations.

Why It Matters

By tightly integrating hardware, data collection, control, and learning, Pixel reduces system mismatches that plague traditional robotics pipelines, enabling more effective and scalable robot learning. This approach lowers barriers to entry for research in mobile manipulation and provides a reproducible, extensible platform for advancing general-purpose robotics.

Key Findings

• Co-designing hardware, data, control, and learning leads to superior performance in mobile manipulation tasks.
• A unified policy learning pipeline effectively transfers skills from human demonstrations to robot execution.
• Hierarchical learning enables language-conditioned, long-horizon tasks requiring dexterity, navigation, and whole-body coordination.