Robotics has under 1% of LLM training data; Khala report sees $3B robot-data spend

0xSammy · x · 2026-10-18

Khala Research's 36-page report "Age of Machines: The Learning Problem" argues robots are still waiting for their ChatGPT moment: language models learned from the internet, while robots lack diverse physical demonstration data.

Key facts and takes

Bessemer estimates only 300K hours of robot manipulation data exist globally, versus a potential need for 100M+ training hours; it projects $3B+ in robot-data spending over two years
Cheap data can be the most valuable: a Georgia Tech/Stanford study found a model trained on 2 hours of teleoperation plus 1 hour of handheld video scored 128 on a task vs 74 for a model trained on 3 hours of robot data; Xiaomi's latest model used 100K hours of handheld gripper footage, aligned on 7,200 hours of robot data, then needed under 10 hours per new task
No consensus on training approach: pixel-to-motor, video-based physics, or massive simulation — NVIDIA is pursuing all three, and all still depend on human demonstration
Open networks use crypto incentives to coordinate simulation, teleoperation and human video collection; Axis reports 7M+ trajectories from 220K+ contributors, lifting LIBERO-Plus task success from 83.9% to 88.8%
AI labs spend billions on human data but almost all of it is text (Meta paid $14.3B for 49% of Scale AI); robots need hand/arm recordings with joint angles and grip dynamics

The report closes with a deep dive on Axis Robotics: strategic positioning, product matrix, academic output and commercial deployments.

Original post →

Robotics has under 1% of LLM training data; Khala report sees $3B robot-data spend

More from Embodied