Latent World Models, Tactile Quadrupeds, and the Real-Time VLA Stack

Tactile-Guided Manipulation on Quadrupeds Without Paired Sensor Data

Teaching quadruped robots to manipulate objects using touch feedback alone, no expensive sensor-motor pairing required. This lowers the data barrier for legged platforms doing real-world manipulation tasks.

Distilling Vision-Language-Action Models to Run at 30Hz for Real-Time Control

Compressing large VLA policies into lightweight controllers that hit real-time control loop speeds. This is the bridge between foundation model capabilities and actual deployment latency.

Multi-Agent Task Allocation via Learned Coordination Without Centralized Control

Decentralized coordination where each agent learns its own allocation policy rather than relying on a central planner. Scales better and handles partial failures, which is what you actually need in multi-robot warehouses.

Policy Adaptation Through Online Fine-Tuning with Minimal Environmental Feedback

Robots that adapt their control policies on the fly using sparse reward signals from their environment. Less reliance on curated datasets, more on real-time learning from deployment.