Modern robots possess impressive physical capabilities. “Their motors operate with remarkable speed and force,” explains Sabrina Neuman, a recent PhD graduate from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL).
However, when facing complex scenarios, particularly human interactions, these mechanical beings often hesitate. “The bottleneck lies within their cognitive processing,” Neuman notes.
Processing environmental stimuli and formulating appropriate responses demands “enormous computational resources,” creating significant delays in reaction time. To address this disconnect between a robot's physical abilities and mental processing, Neuman pioneered an innovative approach called robomorphic computing. This revolutionary method leverages a robot's physical structure and intended functions to create a specialized computer chip that dramatically reduces response latency.
This technological breakthrough promises to transform numerous robotics applications, potentially including frontline healthcare for infectious disease patients. “Imagine robots capable of minimizing exposure risks for both patients and medical staff,” Neuman envisions.
Neuman presented her findings at the International Conference on Architectural Support for Programming Languages and Operating Systems. Her MIT collaborators included graduate student Thomas Bourgeat and Srini Devadas, the Edwin Sibley Webster Professor of Electrical Engineering who also served as Neuman's PhD advisor. Additional contributors comprised Brian Plancher, Thierry Tambe, and Vijay Janapa Reddi from Harvard University, where Neuman now serves as a postdoctoral NSF Computing Innovation Fellow at the School of Engineering and Applied Sciences.
According to Neuman, robot functionality encompasses three critical phases. Initially, perception occurs as the machine gathers data through sensors or cameras. Next comes mapping and localization: “Based on visual input, robots must construct environmental maps and determine their position within them,” she explains. The final stage involves motion planning and control—essentially, determining and executing appropriate actions.
These processes demand considerable time and computational capacity. “For robots to operate safely around humans in dynamic environments, they must process information and respond instantaneously,” emphasizes Plancher. “Existing algorithms cannot execute quickly enough on conventional CPU hardware.”
Neuman acknowledges that while researchers have focused on algorithmic improvements, she believes software solutions alone won't suffice. “What's emerging is the recognition that hardware innovation is equally crucial,” she states. This means evolving beyond standard CPU chips that typically constitute a robot's “brain”—through hardware acceleration.
Hardware acceleration employs specialized hardware components to perform specific computational tasks with greater efficiency. A prominent example is the graphics processing unit (GPU), a chip optimized for parallel processing. These devices excel at graphics rendering because their parallel architecture enables simultaneous processing of thousands of pixels. “A GPU may not be universally superior, but it delivers exceptional performance for its intended applications,” Neuman explains. “You achieve enhanced performance for specific use cases.” Since most robots are designed for particular applications, they could significantly benefit from hardware acceleration—precisely why Neuman's team developed robomorphic computing.
The system generates a customized hardware architecture optimized for a specific robot's computational requirements. Users input a robot's parameters, such as limb configuration and joint movement capabilities. Neuman's system converts these physical attributes into mathematical matrices. These matrices are “sparse,” containing numerous zero values that correspond to physically impossible movements given the robot's anatomical constraints. (Similarly, your arm's movement range is limited by its joints—it cannot bend like an infinitely flexible noodle.)
The system subsequently designs a hardware architecture specialized to perform calculations exclusively on the non-zero values within these matrices. The resulting chip design is therefore customized to maximize computational efficiency for the robot's specific needs. This customization approach demonstrated remarkable benefits during testing.
Hardware architecture created through this method for specific applications significantly outperformed standard CPU and GPU components. Although Neuman's team didn't manufacture a specialized chip from scratch, they programmed a customizable field-programmable gate array (FPGA) chip according to their system's specifications. Despite operating at a reduced clock speed, this chip delivered performance eight times faster than the CPU and 86 times faster than the GPU.
“I was ecstatic about these results,” Neuman admits. “Despite being constrained by lower clock speeds, we achieved superior performance through enhanced efficiency.”
Plancher envisions broad applications for robomorphic computing. “Ideally, we could eventually fabricate custom motion-planning chips for every robot, enabling rapid computation of safe and efficient movements,” he predicts. “I wouldn't be surprised if, two decades from now, every robot incorporates multiple custom computer chips, with this being among them.” Neuman adds that robomorphic computing might enable robots to assume hazardous human responsibilities in various environments, from COVID-19 patient care to heavy object manipulation.
“This research is exciting because it demonstrates how specialized circuit designs can accelerate fundamental aspects of robot control,” remarks Robin Deits, a robotics engineer at Boston Dynamics who wasn't involved in the study. “Software performance is critical in robotics because the real world doesn't pause while robots complete their computations.” He adds that Neuman's advancement could enable robots to process information more rapidly, “unlocking sophisticated behaviors previously deemed computationally unfeasible.”
Neuman's next objective involves fully automating the robomorphic computing system. Users will simply input their robot's parameters, and “the hardware description will emerge automatically. I believe this capability will propel the technology into widespread practical application.”
This research received funding from the National Science Foundation, the Computing Research Agency, the CIFellows Project, and the Defense Advanced Research Projects Agency.
