Bipedal robots like Atlas can walk, balance, and grab. A panel of their builders says scaling them into real workplaces is still unsolved, and they are now naming the deployment problem out loud.
The humanoid robot industry has a visibility problem. It can show you a machine that walks on two legs, balances on shifting floors, and picks up a crate. It cannot yet show you a fleet of those machines earning their keep on a warehouse floor, year after year, outside a controlled demo room.
That gap was the explicit subject of a keynote panel on humanoid robot design at the 2026 Robotics Summit & Expo in Boston, moderated by Mike Oitzman, senior editor at The Robot Report and Automated Warehouse. The panel's most newsworthy thesis came from Alberto Rodriguez, director of robot behavior for Atlas at Boston Dynamics. "Almost all jobs are one of a kind," Rodriguez said, framing the core engineering problem not as raw capability but as variety. His proposed three-pillar deployment strategy — hardware, models and architectures, and integration — treats that variety as the real bottleneck.
A humanoid in this context is a bipedal, mobile robot with arms and hands, designed to operate in spaces built for people. That puts it in a different category from the stationary robotic arms that, as The Robot Report notes, are arguably already mastered for repetitive manufacturing work. Walking and manipulating in an unstructured space is a higher-complexity, unresolved problem, and the panel's job was to map the edges of that problem honestly.
The other panelists were Al Makke, head of humanoid robotics for North America at the parts supplier Schaeffler; Mike Nielsen, chief marketing officer at RealSense, which makes depth-sensing cameras; Aaron Prather, director of the Robotics & Autonomous Systems Program at standards body ASTM International; and Pras Velagapudi, chief technology officer at Agility Robotics. Their roles straddle the full stack: a parts supplier selling into humanoid programs, a sensor vendor, a standards body trying to define test methods, an operator-side integrator, and the robot maker itself.
The honest message, taken together, is that the capability floor has risen. Walking, balancing, and basic manipulation are no longer the open questions they were compared with prior generations of humanoid R&D. The open question is what to do with that capability, at scale, in places where the work changes every hour. Rodriguez's three-pillar framing makes that concrete: hardware has to keep getting more reliable, the models and software architectures that turn sensing into action have to handle a far wider range of tasks, and the integration work of putting a robot into a real workflow has to stop being a research project and start being an operations practice.
The standards work matters here too. ASTM, through the program Prather runs, is one of the bodies trying to define what a working humanoid test should even measure, which is part of why deployment, not just capability, is now the industry's front-burner problem. Until there is a shared way to score how a robot performs in a real workflow, vendors and buyers are arguing past each other.
What to watch: whether the next round of humanoid announcements ships with deployment metrics, the kind of throughput, uptime, and cost-per-task numbers a warehouse manager can plug into a spreadsheet, rather than the dance videos and parkour clips that have defined the genre to date. If the panel's framing holds, that shift is the story of the next two years.