Geek+ Unveils Gravity Control to Correct Vision-Language-Action Policies During Inference
On October 21, 2026, Geek+, a robotics and logistics automation company, released Gravity Control, a causal intervention model designed to correct embodied AI policies at runtime. The model does not require retraining the original policy or adding new demonstration data. Instead, it uses mechanistic interpretability to identify action-control mechanisms inside the model and directly intervenes on relevant latent features during inference.
Gravity Control establishes a closed loop that combines concept interpretability with behavior controllability. According to the company, an operator can use a simple natural language prompt to activate and fine-tune motion features deep inside the model. In out-of-distribution (OOD) manipulation tasks, the architecture raised overall task success from 17.5% to 87.5%, with one specific setting improving from 5% to 95%.
Traditional end-to-end models typically only accept high-level task instructions such as “put the bread on the plate.” When small geometric deviations occur during execution, such instructions cannot convey low-level control parameters like grip depth, wrist rotation angle, or gripper opening. Gravity Control addresses this by building a language-guided action adjustment chain, allowing an operator to give corrective prompts such as “grip deeper and lower when picking up the item.”
Gravity Control has been evaluated on representative VLA architectures including π0.5 and OpenVLA-OFT, the LIBERO simulation benchmark, and physical robot platforms. In real-robot OOD tasks, it achieved zero-shot action recovery without updating original policy weights or adding new demonstration data. The model can also be integrated into higher-level agent or AI-native system suites, enabling an automated detect-diagnose-correct workflow and storing successful corrections as reusable skills for similar future states.
