Robotics
New method enables robots to learn complex tasks in under an hour of real-world training
SLAC uses low-fidelity simulation to pre-train an abstract action space, allowing reinforcement learning robots to acquire sophisticated manipulation skills safely and efficiently.
Researchers have developed a new framework that makes it practical to train sophisticated robotic systems in the real world while minimizing safety risks and sample wastage. The approach, called SLAC, combines simulation-based preparation with real-world learning to help complex machines like mobile manipulators acquire new skills in around an hour.
The challenge of robot learning#
Training robots through trial-and-error (reinforcement learning) poses two fundamental obstacles: the need to explore safely without breaking equipment or harming people, and the requirement to learn efficiently without requiring thousands of costly real-world interactions. Traditional methods that train entirely in simulation often fail when robots encounter the physical world, since simulations rarely capture all the nuances of real materials and forces—a problem known as the reality gap.
How SLAC works#
Rather than training directly on real robots or trying to bridge the simulation gap with high-fidelity physics models, SLAC uses a coarse simulation to pre-train a simplified "language" for robot actions. This intermediate representation—called a latent action space—is designed to emphasize temporal abstraction (grouping related movements together), separate out different control concepts, and naturally align with safer behaviors.
Once this foundation is established, the robot switches to real-world learning using a specialized off-policy algorithm that leverages the pre-trained action space. The approach requires no human demonstrations or manual behavior specifications.
Performance and implications#
When tested on bimanual mobile manipulation tasks—where a robot must coordinate two arms and a base to complete complex jobs—SLAC achieved state-of-the-art results. Remarkably, the system learned contact-heavy whole-body tasks (like grasping and repositioning objects) in under 60 minutes of actual robot operation, substantially faster than competing approaches.
