To me, this suggests that the problem is hard. Last I checked, the state of the art in robotic grasping seems similar to the state of many other AI systems before ML hit the scene. It's super-mathematized all to the questionable end of analyzing how a few points points can optimally apply forces to simple convex polytopes.
A similar feeling exists when you look at the state of path planning for robotic arms. There, collisions must be avoided at all costs because we don't have the mathematics for it. So you make this super precise plan that carefully snakes its way around all the little voxels that happen to become occupied in your occupancy grid. To execute these plans we need to manufacture robots with expensive harmonic gearing and sub-millimeter level repeatability. These types of robots would not be economical for outdoor picking tasks.
To make progress, I think there will have to be new ML techniques and new lower cost robotic hardware developed in tandem.
[1] https://www.appharvest.com/press_release/appharvest-acquires... [2] https://www.therobotreport.com/abundant-robotics-shuts-down-... [3] https://techcrunch.com/2022/02/16/following-acquisition-by-b...
