Robotic manipulation in real-world logistics and manufacturing has long run into an expensive architectural dilemma: end-to-end neural policies require prohibitive, hardware-specific retraining for every new gripper, while classic analytical planners break down under perceptual noise in unstructured clutter. Research published in Nature Machine Intelligence introduces SpaHybGen, a hybrid framework designed to bypass this trade-off entirely by decoupling contact perception from kinematic execution.
Rather than training end-to-end controllers tied to a single manipulator, SpaHybGen trains a neural network to extract spatial contact representations directly from noisy depth data. These hardware-agnostic contact affordances then feed into a differentiable analytical optimizer that solves for stable force closure across arbitrary kinematic hand models. By handling physics analytically and perception neurally, the architecture enables zero-shot deployment across varied mechanical hands without retraining.
In benchmark evaluations, SpaHybGen demonstrated 94.3% to 98.0% grasp success rates across seven distinct robotic hands spanning two to five digits in semi-cluttered environments, alongside maintaining 20 Hz closed-loop dynamic replanning in dense clutter and multi-hand coordination tasks. By open-sourcing the models and codebase, the researchers provide a viable blueprint to compress warehouse and factory automation CAPEX, slashing the engineering overhead of swapping robotic hardware across existing logistics pipelines.