Humanoid robots were the defining hardware story of CES 2026 in Las Vegas this January. Nvidia, AMD and Qualcomm all made robot announcements, and Google DeepMind said it would work with Boston Dynamics on new AI models for the Atlas robot, CNBC reported on January 9, 2026. Fourier's GR-3, a 165 cm care-focused humanoid, also debuted.
What is a robotics platform, exactly?
A platform is a robot built to be programmed rather than to perform one fixed task, and the distinction is architectural. A welding arm on an automotive line executes a pre-planned trajectory thousands of times; a platform ships with sensors, compute and a software stack that third parties extend with new behaviors. Fourier's documentation for GR-3 describes exactly that shape: whole-body movement and balance for scheduled dance demonstrations, real-time board perception and move planning for chess matches against visitors, touch-based interaction and natural conversation, all running on the same hardware in one booth.
The industry's economics explain why every major chipmaker now courts robots. Nvidia CEO Jensen Huang told a CES news conference that the humanoid industry is riding on the work of the AI factories being built for other AI work, meaning the same training infrastructure that serves chatbots also serves machines. Robots give chipmakers a second, physical market for AI compute, and platform builders get pretrained perception and language models they could never afford to train from scratch. That mutual dependency was the subtext of CNBC's CES coverage, which described companies using the show to reveal visions of a future filled with physical artificial intelligence.
How does a humanoid actually decide what to do?
The software stack has three layers, and each was visible at CES in some documented form. Perception comes first: cameras and other sensors feed models that segment the scene and locate objects, as when GR-3 reads a chess board mid-game. Planning comes second: a policy, today usually a learned model rather than hand-written logic, proposes actions that satisfy the task and the robot's physical limits at the same time. Control comes third: joint-level controllers translate plans into torques across the robot's degrees of freedom, and 55 of them means 55 axes to coordinate in exchange for more expressive movement.
- Perceive: onboard sensors and vision models build a machine-readable model of the environment, updated continuously as the scene changes.
- Plan: a learned policy selects actions that advance the task without violating balance, reach or safety limits.
- Act: joint controllers execute the plan, with feedback loops correcting for slippage, contact forces and modeling error.
What makes the current moment different from earlier robot demos is where the intelligence lives. Perception and language models trained on internet-scale data transfer onto robots with modest fine-tuning, so a startup can buy the eyes and the words and concentrate engineering on the body. That division of labor is why startups can field humanoids at a trade show budget that would have required a national program a decade earlier.
Why did care robots lead the demos?
Fourier's positioning of GR-3 as a Care-bot is a market signal, not a sentiment. Per Fourier's announcement, the robot was designed for human-centered scenarios in homes, public spaces and commercial environments, with a soft exterior chosen for approachability. Care environments demand exactly the capabilities platforms are strongest at demonstrating: natural conversation, gentle physical interaction and adaptive behavior around unpredictable humans, whose movements a factory cell would simply fence out.
The timing also fits demographics. First introduced in August 2025 and brought to CES for its American debut, GR-3 arrives in a country where care labor shortages are a documented policy problem, alongside competing demonstrations of laundry-folding assistants and domestic helpers from other makers. The care pitch gives buyers a reason to tolerate the current limits of autonomy, because the alternative in understaffed facilities is often no helper at all.
What can these platforms still not do?
The gap between demonstration and deployment remains the honest story. Chess is a closed problem with a bounded board and a fixed set of legal moves; folding arbitrary laundry, safely navigating a cluttered home, or assisting a frail person out of bed are not closed problems, and none of the CES documentation claims unsupervised long-duration operation in those settings. CNBC's own reporting notes that humanoid home robots remain largely stuck in demo mode despite the hype, a framing that matches what the manufacturers actually claim rather than what the keynote videos imply.
Reliability, safety certification and cost are unsolved at scale, and each compounds the others: a robot that needs supervision delivers less economic value, which keeps unit costs high, which slows the accumulation of deployment data that would improve reliability. Care settings add a certification burden of their own, since a machine that touches people is regulated differently from one bolted to a factory floor. None of these constraints appeared as line items in any CES announcement, and all of them will decide what ships.
What should a platform buyer evaluate?
The CES record suggests a short checklist for anyone comparing platforms on documentation rather than demos. First, the interface surface: a platform is only as open as its published software development kit, and makers that document their SDK honestly list what third parties cannot yet touch. Second, the sensor suite and its replacement cost, because perception hardware fails in the field far more often than actuators. Third, the model update path, meaning whether perception and planning models can be refreshed as new versions ship, or whether the robot is frozen at its factory software version. Fourth, the documented duty cycle, since a robot rated for demonstrations is not rated for eight-hour shifts.
By those criteria, the CES 2026 class shows real progress on capability and openness, and honest gaps on duty cycles and deployment evidence. What the show did document is convergence: the compute, the models and the mechanical engineering have arrived at the same point at the same time. The platform builders are now competing on software ecosystems and application breadth rather than on whether the machines can walk, and the next measurable milestones will be documented deployments with named customers, not dance recitals on a trade show stage.
