Humanoid Robots Are on the Factory Floor — Here's What's Actually Being Deployed

For most of the past decade, humanoid robots appeared at press conferences and keynotes doing carefully choreographed tasks — pouring drinks, doing push-ups, walking carefully on flat surfaces. The implied promise was always a few years away. That promise has started to deliver, selectively and on a smaller scale than the hype suggested, but in ways that reveal which applications actually work in uncontrolled industrial environments.

BMW: Figure's First Real Production Deployment

BMW's Spartanburg plant in South Carolina became the site of one of the first genuine humanoid robot production deployments when Figure 01 units began handling sheet metal transfer tasks in early 2025. The specific task: taking metal body parts from a conveyor, inspecting them visually, and placing them precisely into fixtures for the next manufacturing step. This is exactly the kind of task that humanoid robots are theoretically good at — it requires dexterous manipulation, visual inspection, and precise placement in a space designed for humans — but it is also the kind of task where tolerances matter and errors are expensive.

BMW's public statements have been measured rather than triumphant. The robots are running in limited production areas with extensive monitoring. Cycle times are slower than human workers for most tasks. The value proposition in this phase is not speed — it is learning. BMW is accumulating real-world manipulation data in an authentic industrial environment, data that will be used to retrain and improve the robots over successive generations.

The Figure 01 deployment uses an OpenAI-Figure joint model for high-level reasoning and task planning, with lower-level motor control handled by custom learned policies. The split architecture — a "brain" model for what to do and specialized "muscle" models for how to physically do it — has emerged as a common pattern across humanoid robot developers.

Amazon: Agility Robotics' Digit in the Warehouse

Amazon's relationship with Agility Robotics' Digit robot predates its acquisition of a minority stake in the company. Digit is designed for warehouse environments and has been deployed at Amazon facilities handling tote movement — picking up empty tote containers from one location and transporting them to another. This is deliberately chosen as a starting task: it involves locomotion and manipulation but relatively forgiving tolerances, and the consequences of an error (dropping a tote) are low compared to handling inventory or operating near human workers.

Amazon is explicit that the Digit deployment is a pilot program rather than a scaled rollout. The company operates roughly 75 robotic fulfillment centers with hundreds of thousands of traditional robots, and humanoid robots currently represent a tiny fraction of that fleet. The current goal is operational learning: understanding how bipedal robots navigate real warehouse floors, how they interact with existing infrastructure, and where the failure modes appear in production conditions rather than lab settings.

The economic math for humanoid robots in warehouses is not yet favorable at current prices and reliability levels. Automated storage and retrieval systems (ASRS), traditional robotic arms, and mobile robots do most warehouse tasks more cheaply and reliably than humanoid robots can today. The humanoid case becomes compelling for tasks that require navigating spaces designed for humans that cannot be economically redesigned for specialized robots — loading and unloading trucks, working in mixed human-robot environments, or handling items in locations that traditional fixed-arm robots cannot reach.

Tesla Optimus: The Most Ambitious Deployment

Tesla's Optimus program is the highest-profile humanoid robot effort, and also the hardest to evaluate objectively because Tesla's public communications on the topic mix genuine progress with promotional optimism. What is verifiable: Optimus units are performing tasks inside Tesla's Fremont and Giga Texas facilities, including moving parts between workstations, sorting components, and performing some cable-routing tasks. Tesla has stated that the robots are operating autonomously — without remote human control — on these specific tasks.

The technical approach differs from Figure and Digit in important ways. Tesla trains Optimus primarily using video data collected from its human workers performing the same tasks, using a similar data flywheel strategy to its Autopilot development. The robots watch humans, the model learns the task, and then the robots attempt it — with human oversight and correction in early stages. This approach is ambitious because it tries to generalize across tasks rather than training specialized policies for each specific operation.

Tesla's production targets for Optimus have been revised downward from initial announcements — the company originally projected producing 1,000 Optimus robots in 2024, a number that was quietly walked back. Current estimates suggest a few hundred units are in operation across Tesla facilities, with external sales to automotive and manufacturing partners beginning in limited quantities.

What These Deployments Have in Common

Across BMW, Amazon, and Tesla, several patterns emerge from current humanoid robot deployments:

Structured environments. Every production deployment operates in carefully defined zones with controlled lighting, predictable floor conditions, and specific task parameters. Humanoid robots are not yet navigating genuinely unstructured environments. The apparent flexibility of humanoid form is being used to slot robots into spaces designed for humans, not to navigate arbitrary spaces.

Slow cycle times. Current humanoid robots are slower than trained human workers on every production task. The advantage lies in consistency (no fatigue, no distraction) and in specific applications where human ergonomics make the task unpleasant or unsafe. Hot, heavy, or repetitive tasks that cause injury are better targets than tasks where speed matters.

Data collection as primary value. Every major deployment is being described honestly as a learning phase. The robots are generating real-world manipulation data that would be impossible to collect in a lab — the unpredictable variations in real parts, real lighting, real environments — and that data is the actual product of the current deployment phase.

Human oversight at the task level. In every production deployment, humans are involved in monitoring, error correction, and handling the exceptions that the robots cannot manage. Fully autonomous operation across full work shifts is not the current state.

The 2-3 Year Outlook

The companies best positioned to scale humanoid robot deployment are those with large proprietary training data sets from real-world operations, robust simulation infrastructure for rapid iteration, and existing relationships with manufacturers willing to host continued learning deployments. Tesla, Figure/OpenAI, and Agility/Amazon all have credible positions here.

The tasks likely to see meaningful automation in the 2027-2028 timeframe are those with high economic value and poor human ergonomics: loading and unloading commercial trucks (which causes a high rate of back injuries), repetitive parts handling in automotive plants, and logistics tasks in warehouses that cannot be efficiently restructured for traditional automation. These are not glamorous applications, but they represent billions of dollars in labor costs and significant worker injury risk.

The general-purpose household robot — the one that empties the dishwasher and folds laundry — remains genuinely years away from commercial reality. The unstructured variability of domestic environments, the requirement for safe operation around children and pets, and the need for failure-mode tolerance that does not exist in current systems make this a harder problem than factory deployment by an order of magnitude. The companies that claim otherwise are selling stock, not technology.

What is happening now, in BMW's Spartanburg plant and Amazon's fulfillment centers, is real and meaningful — just not what the press conferences promised.