The Chips Inside Robots

The main kinds of onboard compute

Most robots carry several processors with different jobs. Microcontrollers sit beside the motors and run tight control loops. A main computer, often an embedded system-on-chip that combines CPU cores, a GPU and dedicated AI accelerators, handles perception, planning and learned models. NVIDIA's edge AI computers are one example of this class of board. Some robots add a separate safety controller. FPGAs, chips whose logic can be reconfigured after manufacture, appear where a fixed pipeline must run with very low and predictable delay. Each processor needs power, cooling and software support, so ask a maker for the full list of compute on board rather than only the headline chip.

Why robots need specialised chips

General-purpose processors are flexible but spend energy on work that robots repeat endlessly, such as the physics calculations behind motion planning and control. Researchers at MIT and Harvard demonstrated one alternative, robomorphic computing, which designs hardware around a robot's limb layout and joint structure. Implemented on an FPGA, their design ran eight times faster than a CPU and 86 times faster than a GPU on the target calculations despite a slower clock, MIT News reported. Custom silicon of this kind is still mostly a research topic, but it points towards robot computers shaped to a particular body rather than shrunken versions of data centre machines.

Power, heat and battery life

Every watt spent on computing comes out of the same battery that drives the motors, and the heat has to escape from a closed body. That makes energy per calculation as important as raw speed. MIT's Gleanmer chip, from Vivienne Sze's group, shows how far efficiency can go: it builds 3D maps from compact Gaussian shapes instead of raw pixels, consumes about 6 milliwatts, and needed only about 2.5 percent of the power of the best existing chip for map construction, MIT News reported in 2026. Savings on that scale could mean longer shifts or room for larger models. Ask makers how compute load affects runtime and how the robot copes in hot rooms.

Keeping safety separate from AI compute

The processor running a neural network is a poor place to guarantee safety, because learned models are hard to verify and general-purpose chips are not designed to fail in predictable ways. A safer pattern keeps functions such as protective stops and speed limits on separate, simpler hardware. ISO 13849-1:2023 sets out general principles for designing the safety-related parts of control systems, including their software, and a method for judging how reliably they perform each safety function. When comparing robots, ask which safety functions exist, which controller performs them, what performance level the maker claims, and whether a fault in the AI computer could stop those functions from working.

Sources and further reading

• Designing customized “brains” for robots — MIT News.
  MIT and Harvard's robomorphic computing designs chips around a robot's body; an FPGA version beat a CPU and GPU on speed.
• New chip could help tiny robots traverse complex environments — MIT News.
  MIT's Gleanmer mapping chip runs on about 6 milliwatts, showing how far purpose-built robot silicon can cut power use.
• ISO 13849-1:2023 Safety of machinery — Safety-related parts of control systems — Part 1: General principles for design — International Organization for Standardization (ISO).
  Design principles for control system parts, software included, that carry out safety functions such as protective stops.

Common questions

Can I upgrade the computer in a robot?

Sometimes. Research platforms may expose the main computer and allow replacement, while commercial humanoids usually tie software, cooling and certification to one board. Ask whether an upgrade path exists, whether it voids the warranty and whether the maker's software will run on newer hardware.

Do robots need a GPU to run AI models?

Not always. Small models can run on CPUs or low-power accelerators, and some robots send heavy work to a server. Large vision and language models running onboard usually need a GPU or dedicated AI accelerator to respond quickly enough for real-time use.