MIT researchers have developed a new AI technique to simplify robotic manipulation planning for tasks involving extensive contact points. The approach, called smoothing, reduces complex manipulations down to critical solutions. This enables basic algorithms to rapidly generate effective plans, like lifting heavy boxes.

Humans can leverage their full body to maneuver bulky items, grasping and pressing them against the chest. For robots, each potential contact represents a manipulation variation, resulting in billions of possibilities. Planning becomes intractable.

Smoothing simplifies this by averaging out unimportant intermediate steps, keeping pertinent solutions. It works similarly to reinforcement learning's implicit smoothing. The MIT team created a straightforward model implementing analogous smoothing. This focuses on key object interactions to predict long-term behaviors. Their method matched reinforcement learning's performance for complex plans.

However, finding the remaining post-smoothing solutions can still prove difficult. So the researchers combined their model with an algorithm that quickly iterates through all of a robot's possible choices. This decreased computation time to around one minute on a standard laptop.

The researchers first tested their approach on simulators, assigning robot hands tasks like moving handles, opening doors, and lifting plates. In each instance, the model-based technique equaled reinforcement learning's success, but faster. Tests on real robot hands produced similar results.

While still in early stages, this methodology could enable small mobile robots to manipulate objects using their whole arm or body instead of large grippers. This could cut energy use and costs in factories. The technology may also assist robots exploring other planets, adapting quickly with only onboard computers.

Lead author Terry Suh explains, "If we can leverage the structure of such robotic systems with models, this will speed up the whole decision-making process and plan generation for making contact."

The model relies on a simplified approximation of reality. So it cannot handle dynamic motions like falling objects yet. While effective for slow manipulations, it cannot plan throwing tasks. The researchers aim to improve the methodology to solve highly dynamic problems in the future.

Overall, MIT's manipulation planning research indicates AI smoothing techniques can simplify robotic contact challenges. By focusing on critical solutions, even basic algorithms can rapidly generate manipulation plans once considered unsolvable. This innovative approach may transform future automation technologies.