Diffusion Policies Scale Up, Legged Robots Get Smarter Feet, and Dexterous Hands Learn from Video

Scaling Diffusion Policies with Synthetic Data Augmentation for Robust Manipulation

Synthetic data augmentation is closing the gap on real-world data requirements for diffusion-based manipulation policies. If this holds up in broader testing, it meaningfully reduces the cost of training robust manipulation systems.

Contact-Rich Locomotion via Learned Foot Sensing on Uneven Terrain

This work trains legged robots to use learned foot contact signals for navigating uneven terrain without relying on external perception. Better proprioception is the unsexy unlock that makes real-world deployment actually work.

Learning Dexterous Manipulation from Human Video with Morphology-Agnostic Retargeting

Retargeting human hand video to robot hand morphologies without manual annotation is a long-standing bottleneck and this approach reduces that friction considerably. More usable training signal from cheap video data is a direct win for anyone building dexterous systems.

Hierarchical Task Planning with Foundation Models for Long-Horizon Robot Manipulation

This paper chains foundation model reasoning with low-level robot controllers to handle longer task horizons than single-policy approaches can manage. The framing of when to trust the planner versus the controller is the part worth reading carefully.

Zero-Shot Generalization in Navigation Policies via Semantic Scene Decomposition

Decomposing scenes semantically before running navigation policy inference improves zero-shot transfer across environments without retraining. This is the kind of modular trick that actually ports to new environments without a full data collection cycle.

Unified Representation Learning for Multi-Modal Robot Perception Across Vision and Touch

Fusing vision and tactile signals into a shared representation rather than running separate models reduces inference overhead and improves contact-driven decisions. Practical for anyone building manipulation stacks where latency and sensor fusion both matter.