CES robotics demos tend to look like entertainment until a small detail an unexpectedly careful grasp, a recovery step after a stumble, a hand that seems to “feel” rather than clamp reveals what the builders are actually trying to prove. The weirdness is often the point: it compresses years of mechanical design, control theory, and data collection into a few minutes that a passerby can understand.
One booth reduced that idea to a familiar sound: a ping-pong ball ticking across a table. Sharpa’s humanoid presentation wasn’t only about keeping a rally alive; it was a live argument for dexterity as a platform capability. A Sharpa representative described the system’s 0.02-second reaction time, a number that matters less as trivia than as a proxy for a tight perception-to-motion loop. The underlying claim was that the same loop can be repurposed across tasks an approach reinforced by the company’s emphasis on its mass-produced hand, built around 22 active degrees of freedom and dense tactile sensing.
The contrast on the show floor was immediate: some humanoids were presented as workers-in-waiting, others as spectacle with a hint of unpredictability.
EngineAI’s humanoids drew crowds inside a staged ring, where the machines’ stops, starts, and occasional awkwardness became part of the draw. Even without contact, the choreography highlighted how hard full-body coordination remains when balance, foot placement, and timing all have to close the loop in real time. In parallel, the company’s published specifications gesture at what is changing under the hood: its T800 platform is described as using joint modules capable of 450 Nm peak torque and 14,000 W instantaneous joint power, paired with high-degree-of-freedom structures in the neck, waist, and hands ingredients aimed at the same goal as the crowd-pleasing moves: repeatable, controllable dynamics at human scale.
Unitree, a perennial crowd magnet at CES, leaned into dance again an old trope that still serves a modern engineering purpose. A dance routine is a stress test for actuators, thermal limits, synchronization, and fall recovery, all while the robot remains close enough to people to make safety feel immediate rather than abstract.
Safety is also becoming formalized. A multi-year ISO effort is assembling guidance for humanoids operating around people, including proposed ISO 25785-1, focused on risk assessment and human–robot collaboration. As humanoids move from fenced industrial cells into shared spaces, standards work becomes a design constraint as real as torque or battery density.
Some of the most telling demos were the least flashy. A convenience-store setup from Galbot framed autonomy as workflow: a customer chose an item through an app-like menu, and the robot fetched it from shelves. The value wasn’t speed; it was reliable perception, reach planning, and grasping in a cluttered human environment skills that resemble factory kitting as much as retail.
Laundry, too, keeps resurfacing because it exposes the limits of manipulation. Folding cloth forces robots to handle deformable objects that never present the same geometry twice. Researchers have documented how brittle fixed “pick-and-place” strategies can be, and why adaptive approaches such as AdaFold’s feedback-loop folding perform better when fabric wrinkles or shifts mid-motion. On the floor, Dyna Robotics’ dual arms folding shirts into neat stacks made that research feel operational: a task that looks domestic, but maps directly onto labor-intensive back-of-house work in hotels, gyms, and laundries.
Hovering over all of it was the industry’s push to turn demos into deployments. Boston Dynamics’ new Atlas product pitch centers on repeatability at scale: 56 degrees of freedom, a 2.3 m reach, and the ability to lift 50 kg, along with autonomous battery swapping and integration into industrial software systems. The bigger story is the training loop implied by such machines fleets that learn one task and replicate it across many an idea reinforced by Boston Dynamics’ plan to ship 2026 fleets to Hyundai’s RMAC and Google DeepMind.
CES still rewards weird robots. But the memorable moments increasingly come from something more specific: a glimpse of general-purpose behavior emerging from better hands, tighter control, and data-driven learning plus the safety frameworks that will decide where these machines are allowed to operate once the crowd moves on.