A fundamental belief in robotics is being overturned: the best training data for robots doesn't come from expert demonstrations, but from large crowds producing imperfect, diverse examples. Axis Robotics has released Axis Sim Dataset V1, one of the largest open-source simulation datasets for robot arm manipulation, containing over 50,000 human-teleoperated trajectories across 207 different manipulation tasks. The dataset has already been downloaded more than 160,000 times on Hugging Face, making it the most downloaded open-source simulation dataset for Franka robot arms on the platform.

The release challenges a long-held assumption in robotics that demonstrations must be near-perfect to be useful. Instead, Axis's research suggests that when a large and diverse crowd produces noisy, suboptimal trajectories, the errors average out during training, resulting in more robust policies than those trained on curated expert data. This represents a significant philosophical shift in how the robotics community thinks about data quality.

Why Does Crowdsourced, Messy Data Actually Work Better?

The key insight behind Axis's approach is that data quality exists at the distribution level, not at the individual trajectory level. When thousands of different people perform the same task slightly differently, their uncorrelated mistakes cancel each other out. The resulting policy learns to handle variation and real-world messiness rather than memorizing a single expert's technique.

In benchmarks on a standard robotics evaluation called LIBERO-Plus, continual pretraining on the V1 dataset improved success rates from 83.9% to 88.8% for a baseline model, and outperformed a volume-matched comparison dataset by 37.3%. Importantly, performance improved consistently as the team scaled from using 25% to 100% of the dataset, with no signs of saturation, suggesting that diversity and coverage drive the gains rather than a one-time performance bump.

The largest improvements appeared in scenarios with camera angle changes, sensor noise, and layout variations, exactly the types of perturbations that Axis deliberately introduced during data generation. This suggests the crowdsourced approach naturally prepares robots for real-world deployment where conditions are never identical to training.

How Does Axis Build a Continuously Improving Data Engine?

• Simulation at Scale: Over 200,000 distributed contributors on Axis Hub, a top-3 decentralized application on the Base blockchain, have produced 4.7 million trajectories across 13 different robot embodiments.
• Real-World Egocentric Capture: A managed network of 1,000 or more full-time, quality-control-trained collectors capturing first-person video in real homes and businesses across 14 industries, with 200,000 hours already recorded and growing by 4,000 hours daily.
• Humanoid Loco-Manipulation: Over 500 hours combining mobility and dexterity on real humanoid robots, including Unitree G1 and Booster T2 models, collected through hardware-agnostic teleoperation.
• Human-in-the-Loop Refinement: 500 or more hours of human-gated DAgger post-training, a technique where humans correct robot mistakes on edge cases to improve deployment performance.

What distinguishes Axis from traditional data vendors is the feedback loop. Rather than collecting to a fixed specification and stopping, Axis uses model performance and failure cases to determine what should be collected next, so every training round informs the next collection cycle. Every task and trajectory is recorded on-chain on Base for provenance, and contributors are rewarded for verified work quality.

"The future of Physical AI isn't a static dataset you download once. It's an engine that keeps producing the data the model needs next. Scale gets you broad coverage. Diversity keeps the noise unbiased. The closed loop turns every failure into progress. That's what compounds," said Chris Feng, founder of Axis Robotics.

Chris Feng, Founder of Axis Robotics

The company raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors. Axis was founded by researchers from UC Berkeley, Carnegie Mellon University, Georgia Tech, and Shanghai Jiao Tong University, alongside serial founders who have scaled consumer platforms to over 30 million users.

How Are Robot Companies Using This Data in Practice?

Beyond open-sourcing simulation data, Axis works directly with robot embodiment companies to build customized, embodiment-specific data pipelines and model priors tailored to their hardware. With Booster Robotics, Axis rebuilt Booster's real workspace as a task-aligned digital twin, had distributed contributors collect 42,000 simulation episodes on it, and distilled them into a Booster-specific model prior. With just 30 real-robot demonstrations, that prior reached 87.5% success versus 37.5% for an out-of-the-box baseline model, effectively matching the baseline using half the real-world demonstrations.

Other partners span embodiment companies like Feagine Robotics, model companies including Manycore Tech and Dexmal, and industrial automation firms such as Lotus Cars and Geely Auto. Axis also supplies data to on-chain robotics networks including BitRobot on Solana and OpenRoboto on Bittensor.

The team is already working on V2, which will scale to 1.2 million trajectories across 1,200 tasks, with cross-embodiment generalization and results across multiple vision-language-action (VLA) models, a class of AI systems that combine visual understanding with language and robotic action.

What Other Approaches Are Emerging for Robot Training Data?

While Axis focuses on crowdsourced simulation and real-world egocentric video, other companies are tackling the robot training data problem from different angles. Wetour Robotics has demonstrated Orchestra, a system combining surface electromyography (sEMG) with first-person vision to capture richer human-hand data for robot learning. The approach addresses two key limitations of vision-only systems: cameras cannot directly observe how much force a hand applies, and they lose hand data when the hand is hidden behind an object or leaves the field of view.

In an internal carrying task, Wetour's vision-only pipeline failed to locate the hand in 21.8% of frames, including a longest continuous dropout of 4.32 seconds. The Orchestra system combines an 8-channel sEMG wristband called Conductor with a first-person camera called VisionLink, designed to fuse both streams into one synchronized record of movement, effort, and action timing. The cross-modal correction strategy using muscle activity during missing visual intervals remains in validation.

"Physical AI needs more than video. A camera can show where a hand moved, but not how hard it worked, and it can go blind at the exact moment contact happens. Orchestra is designed to add force-related information and continuity to human demonstration data," stated Nan Zheng, Chief Executive Officer of Wetour Robotics.

Nan Zheng, Chief Executive Officer of Wetour Robotics

Wetour's demonstration videos show five real-world tasks combining precision, changing hand effort, and frequent visual occlusion, including packing a lunch box, sorting pills, disassembling a pen, measuring a drone with calipers, and installing a drone propeller. The company believes adding force-related data and greater resilience to visual occlusion can make human demonstration data more useful for fine manipulation, compliant control, and future human-robot collaboration.

Both approaches reflect a broader recognition in the physical AI community that the quality and richness of human demonstration data directly determines how well robots can learn to perform real-world tasks. Whether through crowdsourced diversity or multimodal sensor fusion, companies are moving beyond simple video capture to build more complete records of human action, effort, and intent.