Figure AI has crossed 86,000 weekly active users contributing to its robotics dataset, according to CEO Brett Adcock, marking a significant milestone in the race to build the training data foundation for humanoid robots. However, the company's claim to have "the largest and most diverse robotics dataset in the world" remains unverified by independent benchmarks or direct comparisons to competitors.

Why Dataset Size Alone Doesn't Guarantee Success in Robot Training?

For robotics teams developing embodied AI systems, raw user count tells only part of the story. The actual usefulness of a dataset depends on factors that Figure has not publicly disclosed, including what types of tasks are represented, how the data is labeled, which hardware platforms are covered, and whether the uploaded information can actually be traced back to specific robot actions. A dataset with millions of pixels of video footage might sound impressive, but if those pixels don't correlate to meaningful robotic behaviors, the collection becomes less valuable for training.

Figure's announcement included a live map viewer displaying each pixel of data as it uploads globally, creating a visual impression of scale and geographic diversity. Yet the company provided no figures for total data volume, task coverage, hardware coverage, or geographic distribution. These metrics would be essential for robotics researchers evaluating whether Figure's dataset can actually support the diverse training needs of different humanoid designs and industrial applications.

How Are Competitors Approaching the Robot Data Problem?

The robotics industry is fracturing into competing philosophies about what kind of data should teach robots how to manipulate objects. This divergence reflects a fundamental debate about the path to embodied AI:

• Web-Scale Video Models: Companies like those developing Dyna-2 ingest internet footage to train visual understanding, betting that large-scale passive video can transfer to robotic control.
• Fleet Teleoperation: Enterprise robotics teams collect data by having human operators remotely control robots, creating task-specific training signals but potentially introducing unnatural movement patterns.
• Direct Human Demonstration: Newer entrants like Reward AI argue that robots should learn from natural human movement captured through specialized hardware, bypassing both teleoperation and on-robot training entirely.

Reward AI, which exited stealth on September 14, 2026, is challenging the conventional wisdom by claiming that teleoperation rigs and parallel-jaw grippers systematically corrupt training data. The startup argues that traditional teleoperation forces operators into unnatural, sluggish movements, filtering out the subconscious reflexes and tactile corrections that humans use naturally when manipulating objects.

To capture those nuances, Reward AI engineered the Omnibody Hand, a wearable data-collection glove with a compact 7-degree-of-freedom functional architecture. Rather than trying to mirror every biological joint with a fragile 20-plus degree-of-freedom copy, the device focuses on precision and power grasping through independent thumb and index finger flexion, while the middle, ring, and little fingers move together to wrap around objects. The glove uses electromagnetic tracking alongside visual-inertial sensing to achieve sub-millimeter accuracy, reducing mean trajectory error from 24.9 millimeters to 9.5 millimeters at peak velocity.

Reward AI's foundation model, called OM-1, claims to generalize zero-shot across tabletop arms, heavy industrial manipulators, and humanoids, learning novel long-horizon skills from less than 30 minutes of demonstration data. The company states that OM-1 uses no teleoperation and zero on-robot physical data in its pretraining corpus, instead consuming synchronized multimodal streams of visual context, proximity sensing, tactile contact, and force trajectories.

What Real-World Tasks Can These Systems Actually Perform?

Reward AI demonstrated OM-1 executing complex coordination tasks at full human speed, including unplugging an RJ-45 Ethernet cable, a deceptively difficult manipulation challenge requiring continuous tactile feedback to depress a locking tab before applying pulling force. The system also performed rapid household workflows like bimanual laundry folding, bartending, and packaging consumer electronics within sub-30-second task cycles, while exhibiting adversarial error recovery when objects were perturbed or pulled away.

To bridge the gap between human-level policy planning and real robot hardware, OM-1 decouples cognitive planning from motor actuation. While the transformer-based policy evaluates visual and tactile histories to predict action chunks, an underlying control layer runs on an independent high-frequency clock trained via reinforcement learning in simulation. This controller learns to absorb system delays, unmodeled motor backlash, and sudden physical loads without waiting for the next inference frame from the main model.

How Is Europe's Humanoid Sector Reshaping the Competition?

While Figure AI builds its dataset empire, European robotics ventures are emerging with different technical approaches. A Munich stealth startup led by Nikolai Ensslen, co-founder of motion-control specialist Synapticon, has revealed its alpha development platform called Kyle and plans to unveil its first-generation product in early 2027. The team includes Oliver Groth, a former Google DeepMind research scientist, and Boris Belousov, an AI engineer who previously held a senior research role at DFKI's Systems AI for Robot Learning department.

The Kyle prototype demonstrates what Ensslen describes as "one of the better gaits" among humanoids he has observed, despite its boxy, unfinished appearance. The rectangular frame reflects the team's rapid prototyping approach, and the company deliberately showed the development platform without blurring it for the first time on September 14, 2026, to signal progress toward Generation 1. Ensslen credited Martin Riedmiller and the team for the walking performance, noting that the distinction between the development mule and the finished product matters for understanding the engineering process.

"I have watched a lot of humanoids walk over the last years. In my honest assessment, this is one of the better gaits out there, and it is still our very early stage," said Nikolai Ensslen.

Nikolai Ensslen, Co-founder and CEO, Munich Stealth Startup

The Munich team's move into humanoid development signals a broader European push into physical AI. The company has emphasized that it owns both the hardware and AI stack, building Kyle from scratch in a few months with a focus on reinforcement learning as the finishing ingredient for robot control. The public introduction serves partly as a recruitment message, with Ensslen inviting people who have built humanoids or shipped hardware at scale to join the venture.

Figure AI's dataset milestone reflects the industry's recognition that data will be as critical as silicon and motors in the race to deploy humanoid robots at scale. Yet the company's claim to leadership remains unsubstantiated by the metrics that actually matter to robotics researchers: coverage, labeling, provenance, and the relationship between uploaded pixels and robot actions. As competitors like Reward AI and European startups pursue alternative data collection and training philosophies, the question of which approach will ultimately produce the most capable robots remains wide open.