Unverified60% confidenceFactExact time
It is estimated that all publicly available robot manipulation data worldwide amounts to fewer than 1 million trajectories in total, a gap of several orders of magnitude compared to language model training data.
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
8/20/2026
First Seen
Valid until: 11/18/2026
Sources
Related Claims
UnverifiedCompanies such as NVIDIA have in recent years heavily invested in simulation training and synthetic data, having robots practice millions of times in virtual environments to compensate for the lack of real-world data.68% similarUnverifiedThe team expects thousands of people to train robot behavior policies on real hardware.68% similarUnverifiedThe supply bottleneck in training data is an unavoidable threshold for NVIDIA, Tesla, and humanoid robot manufacturers.67% similarUnverifiedTeleoperation data is currently recognized as the highest-quality robot training data, where operators control real robots in real time via gamepads, VR devices, or master-slave robotic arms, with almost no transfer loss.65% similarUnverifiedThe robots.txt protocol relies on voluntary compliance, has no legal binding force, and was originally designed without anticipating the scenario of large-scale language model training.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/775873API
curl https://kongchang.com/api/v1/knowledge/claims/775873MCP
get_claim(id=775873)