Jev reads the scene state and chooses arm movements and gripper commands.
What We Are Testing
Can Jev turn spatial understanding into effective robot control, without a predefined grasping skill?
Short-Term Goal
Approach, align, grasp and lift an object, then extend to transport.
Long-Term Goal
Shuqing's vision: a harness for agents to observe, act, verify and adjust in a dynamic world. We test high-frequency robot control with Jev and explore System 1 / System 2 collaboration with GPT-6 Astra for planning and execution.
Current Progress
Conclusions & Next Steps
Metadata alone is not enough in our tests; real-world ground truth is also hard to obtain. We see a need for multimodal input, which Jev does not support.
Next: test multimodal input with the OpenAI Decisions API on tasks of the same difficulty.