CAIP study evaluates dual Sharpa Wave hands across six manipulation tasks
An external Berkeley-NVIDIA study evaluated action-image pretraining on a Dexmate Vega platform with two Sharpa Wave hands, reaching 76% average success across six real-world tasks.
Evidence notes
- The setup paired Dexmate Vega arms with two 22-DoF Wave hands, a ZED X Mini head camera and two ZED X One wrist cameras.
- CAIP used 200 demonstrations per task, with 150 for pouring, and exceeded the strongest reported baseline by more than 30 percentage points on average.
Company context
Founded by leadership from Hesai Technology, with operations across Singapore (R&D), Shanghai (manufacturing), and Mountain View (business). Focuses on embodied AI systems combining dexterous manipulation hardware with model-driven control. Develops both robotic platforms and in-house AI models for long-horizon task execution.