Agility Robotics, maker of the bipedal Digit humanoid, says it is exploring a range of designs and form factors, including wheeled variants, to fit different operating environments and customer needs, according to The Robot Report. The disclosure matters less as product news than as a signal: the company most associated with legged humanoids is hedging on legs.
Wheels alongside legs, not instead of them
The Robot Report's reporting does not describe a wheeled Digit as imminent, nor does it specify which facilities or tasks would use one. It frames form factor as a portfolio question rather than a single bet. Digit is the platform most associated with legged humanoid design, so a company hedging toward wheels for some deployments is treating legs as one tool among several, not a fixed identity.
Buyers comparing platforms on the humanoid spec sheet should not assume "humanoid" implies one fixed body plan. A warehouse with flat, uncluttered floors may suit a wheeled base better than legs built for terrain it will never encounter. Evaluate the operating environment first, then ask which form factor a vendor can actually ship into it, rather than assuming a legged robot is the default.
A general model driving the robot, not a bespoke one
The next development comes from research, not industry. According to The Decoder, researchers from Stanford and Caltech had a humanoid robot powered by GPT-6 Astra independently tidy up an unfamiliar kitchen. The system, called HomeBody, skips a specially trained control layer: the language model itself guides the robot's actions, rather than handing off to a controller trained for that specific task and environment.
That distinction matters more than the kitchen-tidying itself. Most humanoid demonstrations rely on a control policy trained for a narrow task in a known environment, often after extensive teleoperation or simulation. A general-purpose model driving a robot in an unfamiliar space, without that bespoke layer, points toward software decoupled from specific hardware and tasks. This is a research result, not a shipping product. It says nothing about reliability, speed or safety margins outside a lab setting.
The software layer is where the real competition may sit
Other reporting points the same direction. The Robot Report describes General Robotics as composing specialized capabilities rather than training a single model to do everything, aiming to help robots adapt across tasks and form factors. Robohub reports that Intrinsic, the AI robotics group at Google, is open-sourcing parts of its physical AI platform. None of these approaches matches HomeBody's single-model control exactly, but all three treat the control stack, more than the chassis underneath it, as the layer that determines what a robot can do.
If that holds, evaluating a humanoid platform becomes two separate decisions: which body suits the deployment environment, and which control approach (proprietary, composed or general-purpose model) actually gets tasks done reliably. Hardware specs and price, tracked on our humanoid price trackeranswer the first question. They say very little about the second.
Context: a mature industry, an unsettled safety debate
The IFR reports 5 million robots are now at work in factories worldwide, with China leading installations and the US and Germany also significant markets, per The Robot Report. Industrial automation at that scale runs on narrow, well-tested control software bolted to fixed hardware, the opposite of what HomeBody demonstrates. General-purpose humanoids steered by large language models remain a research-stage idea layered on top of an industry that mostly still works the old way.
The Robot Report also cites industry observers saying Asimov's fictional Three Laws of Robotics are not enough to keep robotics and AI safe, a debate sharpened precisely when control shifts from a narrow, auditable policy to a general model making its own decisions in someone's kitchen or warehouse. Buyers evaluating any platform that leans on general-purpose AI for control should ask vendors what testing and constraints govern that model's behavior, not just what it can do in a demo.
What this means for a buying decision today
- Don't lock into a form factor on the assumption it is fixed. Agility Robotics' own signal is that legs versus wheels is a deployment-specific choice, not a philosophy.
- Ask what drives the control software. A bespoke trained policy, a composed set of specialized capabilities and a general-purpose model steering the robot directly all fail differently. HomeBody is a lab result, not a purchasing recommendation.
- Weigh the software roadmap alongside current hardware specs. If control software is where vendors differentiate next, a platform's upgrade path may matter as much as its current spec sheet.
None of this is a reason to delay a pilot that already fits a well-defined task. It is a reason to ask any vendor, Agility Robotics included, what form factors and control approaches they will support over the life of a contract, not just what they can ship now. Coverage of these shifts continues on our robotics news page.