Accelerated Low-power Motion Planning for Real-time Interactive Autonomy

The inability to plan in real-time is a major reason why robot manufacturing is restricted to high-value, high-volume products, in carefully engineered factories where robots blindly repeat pre-programmed trajectories. Today, robots simply cannot generate collision-free motion on the fly. We are solving this core problem—real-time motion planning in unstructured, fluid environments and on diverse Size, Weight, and Power (SWaP)-constrained robot platforms—by co-designing custom motion planning hardware with novel planning algorithms. Just like the GPUs that revolutionized graphics, our motion planning processors will exploit the phenomenal speed and massive parallelism available to dedicated circuitry to make thousands of plans per second. Real-time motion planning is only possible with the combination of specialized hardware and novel algorithms, and we are the first to develop it. Our processors will generate thousands of plans per second, which is faster than perception and enables the robot to create multiple viable motion plans and choose the one that is most convenient (e.g., for grasping) or least likely to collide with dynamic obstacles with uncertain trajectories.