COBALT Launches Smartphone Teleoperation Platform For Robot Data Collection
Efficiency enables everyone! COBALT's modular, cloud-first architecture supports concurrent, global teleoperation via smartphones, VR, or 3D mice.
To maximize efficiency, we utilize vectorization to support multiple parallel users on a single GPU.The result? A near 4x reduction in data collection costs per 1,000 demos on an entry-level NVIDIA T4.
Animesh Garg@animesh_gargRobotics is still data starved. Collecting high-quality robot demonstrations remains brutally slow and expensive. Introducing COBALT: A cloud-native teleoperation platform designed for large-scale robot learning. We are democratizing data collection by leveraging the hardware everyone already owns: the smartphone All you need is to download an app (today)! Read on for more!
Scale should match quality! Our user study shows that smartphone control yielded shorter completion times and smoother trajectories than traditional inputs (VR, SpaceMouse), with the lowest calibration errors.
Real-time metric filtering ensures every collected intervention is rigorous.
Animesh Garg@animesh_gargEfficiency enables everyone! COBALT's modular, cloud-first architecture supports concurrent, global teleoperation via smartphones, VR, or 3D mice. To maximize efficiency, we utilize vectorization to support multiple parallel users on a single GPU.The result? A near 4x reduction in data collection costs per 1,000 demos on an entry-level NVIDIA T4.
In a pilot study, we crowdsourced >7,500 demonstrations (>50 hours of data) in just 5 days across 9 countries, purely using smartphones.
We validated this dataset by training SOTA algorithms (ACT, DP) and confirmed compatibility with real physical robots (Franka Panda, YAM arms)
We did all this without burning VC Money :😉
Animesh Garg@animesh_gargScale should match quality! Our user study shows that smartphone control yielded shorter completion times and smoother trajectories than traditional inputs (VR, SpaceMouse), with the lowest calibration errors. Real-time metric filtering ensures every collected intervention is rigorous.
Try out COBALT today!
Arxiv: https://arxiv.org/abs/2605.19138v1 Website: http://cobalt-teleop.github.io Code: https://github.com/pairlab/COBALT
iOS App: https://apps.apple.com/us/app/cobalt-teleoperation/id6743640565 Android App: released soon.
co-led by @agarwal_ayush05 & @angandhi04 alongside collaboration with @jerthesquare_, @omarrayyann, @sarswat_aryan, Ranjani Koushik, @MasoudMoghani, @AjayMandlekar and @animesh_garg.
Animesh Garg@animesh_gargIn a pilot study, we crowdsourced >7,500 demonstrations (>50 hours of data) in just 5 days across 9 countries, purely using smartphones. We validated this dataset by training SOTA algorithms (ACT, DP) and confirmed compatibility with real physical robots (Franka Panda, YAM arms) We did all this without burning VC Money :😉