Technology & Humanity
Robots Gain 'muscle Memory' For Planning Moves In Milliseconds
From Phys.org Technology · November 11, 2026
Researchers from MIT's CSAIL and Carnegie Mellon University's MLD have developed a new approach called "Workspace Models" to improve robotic responsiveness. According to the study published on arXiv, the system uses large vision-language models (VLMs) to perform reasoning offline, which then trains a smaller, lightweight module. This process effectively transfers a form of "muscle memory" to the robot, allowing it to compress historical observations into latent vectors for rapid decision-making.
Lead author Nitish Dashora stated that this approach provides a significant latency upgrade over previous agentic methods. In experiments involving a robotic arm, the system successfully performed tasks like picking up and putting away cubes approximately 92% of the time. Researchers noted this was roughly 25 percentage points higher than the next-best approach. The researchers suggest this method could help robots function more reliably in dynamic environments, such as kitchens, by better distinguishing which past details are relevant for future actions.
A factual summary prepared by The Quantum Disciple. The original report belongs to its publisher.