COBALT: Crowdsourcing Robot Learning via Smartphone Teleoperation

Researchers have developed COBALT, a cloud-based teleoperation platform that enables crowdsourced robot learning using smartphones. The system allows multiple users to simultaneously control robots in simulation and the real world, addressing the scarcity of large-scale demonstration data for imitation learning. By leveraging vectorized environments and load-balanced infrastructure, COBALT supports concurrent teleoperation on a single GPU, significantly reducing costs. Operators can connect from anywhere using devices like smartphones, VR headsets, 3D mice, or keyboards. An in-memory data cache and efficient video streaming maintain synchronization at 20 Hz with sub-100 ms latency for up to 8 concurrent users per GPU. The system also demonstrates stable operation with 256 simulated clients across 8 GPUs. This work aims to democratize robot learning at scale.