The humanoid robot deployment race looked radically different six months ago. Figure AI was showing off impressive warehouse footage. Boston Dynamics had Atlas doing parkour. Tesla's Optimus was being tested at select facilities. The narrative was simple: humanoid robots are coming to manufacturing in 2026.
The reality? It's messier, slower, and far more interesting than the headlines suggested.
Who's Actually Deployed, Where, and Doing What
Figure AI remains the furthest along in production deployment. By late 2025, Figure had units operating at BMW manufacturing facilities in Germany, handling repetitive assembly tasks in controlled environments. But—and this is crucial—these aren't doing full-spectrum factory work. They're operating in structured contexts with minimal variation: picking specific parts, placing them in precise locations, following predetermined sequences. The tasks work because the environment is engineered around the robot, not the other way around.
Figure's Gen-2 humanoid has made genuine progress on gripper dexterity and proprioceptive feedback. But deployment is still heavily scripted. Each new task requires weeks of demonstration and fine-tuning. The company has been honest about this: true generalization remains years away.
1X Technologies (formerly Halodi Robotics) has deployed their EVE humanoid in industrial settings across Scandinavia and Germany. The 173cm tall unit handles material handling and basic assembly. Like Figure, 1X operates in what they call "structured flexibility"—environments that have variation, but bounded variation. The robot can handle some environmental differences, but not unstructured chaos.
Boston Dynamics took a different approach. While Atlas continues to impress in research demonstrations—that viral parkour video from late 2025 was real—Boston Dynamics has been deliberately cautious about production claims. Atlas is being used in limited factory pilots, primarily for R&D validation rather than production deployment. This restraint actually increases credibility.
Agility Robotics (Digit) and Apptronik (Apollo) have units in testing phases at various facilities, but neither is reporting significant production deployment yet. Apptronik has been more transparent: they're targeting warehouse and logistics work where tasks are highly repetitive and environmental variation is minimal.
The Three Core Failures That Still Exist
The gap between demo and deployment reveals three persistent limitations:
Unstructured Environmental Variability
A factory floor looks clean in marketing footage. In reality, it's cluttered with unexpected objects, workers, equipment that moves. Humanoid robots struggle dramatically with objects outside their training distribution. A box rotated 15 degrees. A part placed slightly differently. A new worker in their work envelope. These things cause failures that cascade—the robot stops, requires human intervention, needs reprogramming.
This is why deployment success correlates directly with environmental control. BMW didn't deploy Figure into their busiest, most chaotic section. They engineered a structured zone around it.
Cost Remains Prohibitive
Figure's hardware costs around $150,000 per unit. 1X's EVE is in a similar price range. A warehouse worker costs roughly $35,000–50,000 annually all-in. Even with five-year deployment horizons, the unit economics only work for the most expensive, dangerous, or physically demanding tasks. And that's assuming zero downtime and perfect utilization.
Reliability and Mean Time Between Failures
Figure's units are still experiencing failure modes that require technician intervention. Motor failures, sensor miscalibration, software edge cases that didn't appear in training. MTBF (Mean Time Between Failures) is improving but remains measured in days or weeks for extended operation, not months. For context, industrial machinery typically targets 99.5%+ uptime.
What 2026 Actually Looks Like
There are roughly 50–60 humanoid robots deployed in active production globally as of March 2026. This is genuinely up from under 10 eighteen months ago. It's real progress. But it's also orders of magnitude smaller than early 2024 speculation suggested.
Goldman Sachs revised their 2030 deployment forecast downward in their 2026 robotics report, citing reliability challenges and limited task scope. Most analysts now project that by 2030, humanoid robots will occupy a specific niche: high-cost labor markets, dangerous environments, and tasks with significant variance that still operate within engineered boundaries.
The real story of 2026 is not that humanoid robots have arrived. It's that the companies willing to be honest about limitations—Boston Dynamics, early-stage players investing in silicon rather than just hardware—are building credibility for what's actually possible in 2029–2031.
The Ones to Watch
Figure AI remains the deployment leader by volume, but watch where they expand. Each new facility tells you what their actual capability frontier is. 1X's approach of building environments first, then deploying, may prove more scalable than Figure's tighter coupling to specific manufacturers. And Boston Dynamics' continued research focus suggests they're building something more general—but they're not going to announce it until it's actually ready.
The 2026 lesson: the robots are real, the deployments are real, but the hype cycle was off by roughly 3–5 years. That's progress worth celebrating, even if it doesn't fit the 2024 narrative.