Realtime Robotics is linked to motion-planning research that reported a three-orders-of-magnitude speedup using an FPGA prototype. That historical result is not a verified specification for a deployed ASIC matching this article’s headline. The useful distinction is between collision-checking latency, complete planning time, and production-cell performance.
What evidence supports the 1000x motion-planning claim?
The 2016 Duke University paper describes a field-programmable gate array, or FPGA, prototype and reports a three-orders-of-magnitude performance improvement through hardware/software co-design.
IEEE Spectrum’s December 2016 account identifies Realtime Robotics as the commercialization startup and describes a future move from FPGAs to application-specific integrated circuits, rather than reporting a deployed production ASIC benchmark.
How did the motion-planning accelerator work?
The research accelerator combines aggressive precomputation of collision data with parallel collision-detection circuitry, rather than simply accelerating conventional geometric calculations unchanged.
In IEEE Spectrum’s explanation, a probabilistic roadmap represents possible robot motions as edges, and each collision-detection circuit checks one edge against incoming obstacle locations.
The prototype described in the paper was interfaced to a Jaco2 robotic arm in the researchers’ laboratory for pick-and-place demonstrations.
How fast were collision detection and complete planning?
The paper reports less than 650 microseconds from sending obstacle data to having a motion plan, with 16 microseconds spent on collision detection in the described experiment.
For a 2,500-edge roadmap, the paper compares custom FPGA collision-detection hardware against CPU and GPU hashset implementations using the same aggressive precomputation.
| Implementation | Reported time | Scope |
|---|---|---|
| Custom FPGA hardware | 16 microseconds | Collision detection, 2,500-edge roadmap |
| CPU precomputed hashsets | 9,550 microseconds | Collision detection, 2,500-edge roadmap |
| GPU precomputed hashsets | 1,100 microseconds | Collision detection, 2,500-edge roadmap |
The authors identify operating-system delays and graph search as major contributors to complete planning time after collision detection is accelerated.
What commercial RapidPlan hardware was reported?
The Robot Report reported in April 2019 that Realtime Robotics was shipping RapidPlan, combining a hardware motion-planning accelerator with a software roadmap-generation toolkit.
That report attributes to the company the ability to evaluate up to 800,000 stored motions at 30 frames per second; it does not present the same measurement as a complete planning latency per production cell.
How do RapidPlan and Resolver differ?
Realtime Robotics currently describes Resolver as cloud-based preplanning of collision-free paths and interlocks, while positioning RapidPlan for systems that must operate in changing environments.
What is still needed to verify the ASIC headline?
[VERIFY: Which deployed Realtime Robotics ASIC model, dated datasheet and reproducible benchmark establish the claimed 1000x complete motion-planning speedup, including the baseline hardware, robot count and workload?]
About the Author
