Technical Demonstrations
A closer look at the tech behind our launch.
Real-world Data Collection
We have access to diverse environments for real-world data collection where dexterous manipulation is key, including in expert settings.
Vision vs Multimodal labelling
A comparison of video-based feature extraction (including HOPE-estimated contact and pressure) versus multimodal effort estimation, showing how adding EMG and tactile sensing improves effort and contact awareness.
Force and Contact Labeling
An example data segment of pick and place task with objects of various weights. EMG derived relative force estimates shown on the right, where heavier objects shows higher grip intesity. This data will be crucial for training force-aware manipulation policies.
Robot Training
A comparison between policies that are trained using vision alone versus using vision + EMG-derived features. The EMG-based model shows greater force and contact awareness.
Dexterous Hand
We have a full stack dexterous hand and robotic arm setup to ensure force and
contact data actually improves manipulation performance. We believe that dexterous
hands are the form factor that best leverages EMG derived data. Yes, this demo is
tele-op-ed, but the autonomous version is coming soon!
Thanks to Proception and
Gento Robotics for their amazing hardware!