Tesla Optimus, Figure 02, and Agility Digit are all in production deployments. We compare capabilities and costs.
The humanoid robot has quietly stopped being a demo-day spectacle and become a line item on a factory floor budget, and the companies that got there first are now competing on uptime and accuracy rather than viral video clips.
Three companies, three different bets
By the first quarter of 2026, Tesla, Figure AI, and Agility Robotics all have humanoid or human-scale robots performing genuinely productive work in commercial settings rather than staged pilots. Each company has made a distinct bet about where these machines add the most value, and the divergence is becoming one of the more interesting strategic splits in robotics.
Tesla's Optimus Gen 3 is the most widely deployed by raw unit count, with thousands of units now working inside Tesla's own manufacturing facilities on tasks like wiring harness insertion, battery cell placement, and visual inspection. Running at roughly sixty percent of human speed but with accuracy above ninety-nine percent and no fatigue-related error creep over a shift, Optimus is essentially a vertically integrated bet: build the robot to serve your own factories first, and treat that internal deployment as the proving ground before selling externally.
Logistics and mixed human-robot spaces
Figure AI has taken the opposite path, aiming squarely at logistics and warehousing rather than in-house manufacturing. Its Figure 02 units are performing bin picking, parts transport, and quality checks inside automotive plants, and the model behind them is built to understand natural-language instructions from floor supervisors directly, cutting down on the pre-programmed, task-specific routines that older industrial robots required. That flexibility matters most in environments where the task mix changes daily and reprogramming a traditional robot arm for every new job would be impractical.
Agility Robotics has carved out a third niche: environments built for human bodies that traditional industrial robots handle poorly, narrow aisles, stairs, and workspaces where people and machines operate side by side. Its Digit robot is being used for tote movement and shelf restocking inside fulfilment centres, leaning on its bipedal form factor specifically because it can navigate spaces designed around human dimensions without any facility redesign.
What unites all three deployments is that none of them required redesigning the facility around the robot, which was the fatal flaw in most previous generations of warehouse and factory automation. Traditional industrial robots demanded caged work cells, fixed mounting points, and carefully choreographed workflows; humanoid and bipedal robots are explicitly designed to slot into spaces and workflows that were built for people, which is a large part of why deployment has moved this fast.
The economics, honestly assessed
The cost math is compelling but not yet a slam dunk. A humanoid robot currently runs somewhere between fifty thousand and a hundred fifty thousand dollars and can realistically replace roughly half to eight-tenths of a full-time equivalent worker depending on the task, since these machines are still slower and less dexterous than an experienced human at most jobs. At today's prices, that works out to a payback period of eighteen to twenty-four months for operations running three shifts, which is attractive for large-scale manufacturers but a harder sell for smaller operations with tighter capital budgets or single-shift schedules.
What is driving adoption despite that uneven payback is not pure cost savings but reliability and flexibility: a humanoid robot does not call in sick, does not need retraining when reassigned to a new task the way a purpose-built machine would, and its error rate does not creep upward at the end of a long shift the way a tired human worker's might.
Where this goes next
As unit costs fall and the underlying foundation models improve, the economic case is likely to strengthen quickly rather than gradually, since most of today's cost is manufacturing scale rather than fundamental technology limits. The more interesting competitive question over the next two years may not be whose robot is more dexterous, but whose foundation model generalises best across the widest range of tasks without needing a specialist retraining cycle for every new job on the floor.
For teams trying to evaluate the underlying AI rather than the hardware, Vincony.com's Deep Research tool can pull together benchmark comparisons across the robotics foundation models powering these deployments, from RT-3 to GR00T, synthesising the fragmented research literature into something a procurement or engineering team can actually act on.