Adaptive Policies, Federated Fleets, and the LLM Reward Engine

Unpaired Tactile Video Learning for Contact-Rich Manipulation

This paper introduces a way to teach robots contact-rich manipulation by learning directly from tactile video without needing paired action data. For builders working on assembly or dexterous tasks, this is a real unlock because labeled manipulation data is the bottleneck that makes these pipelines expensive.

Rapid Online Morphology Adaptation for Legged Robots

The paper tackles adapting robot policies when the hardware morphs mid-mission, think broken links or changed payloads. This is the kind of robustness work that makes real deployments less painful because you stop throwing away the whole policy when a leg gets snagged.

Generative Latent Dynamics for Sample-Efficient Robot Learning

Generative latent dynamics lets you learn a compact world model and plan directly inside it, which cuts the sample complexity for long-horizon tasks. Builders deploying mobile manipulation should watch this because it compresses the planning loop toward real-time.

Geometry-Aware Novel View Synthesis for Robotic Scene Understanding

Novel view synthesis applied to robot perception gives you geometrically consistent scene understanding from sparse viewpoints, cutting sensor cost. Stack this with your VLA pipeline and you get richer spatial grounding without buying more cameras.

Diffusion-Latent Bimanual Rehearsal for Coordinated Manipulation

Diffusion-based bimanual rehearsal lets two arms practice coordinated manipulation in learned latent space before acting in the real world. If you are building dual-arm cells this is the kind of data augmentation that collapses your demo-to-deployment gap.

Federated Policy Learning Across Quadruped Fleets

Federated learning across a fleet of quadrupeds means each robot contributes experience without shipping raw data to a central server. For teams running swarms in sensitive environments this preserves privacy while still scaling your policy library.

LLM-Driven Reward Shaping automates RL Task Specification

Using language models to automatically generate and critique reward functions for RL is the new meta that saves weeks of reward engineering per task. Pair this with your sim stack and you have a pipeline that ships novel behaviors faster than hand-tuning ever could.

4D Topological Planning for Dynamic Multi-Robot Coordination

Topology-aware 4D planning adds time as an explicit dimension so the robot reasons about dynamic obstacles and coordination sequences. This matters for warehouse and logistics bots where the scene never sits still and replanning at 1 Hz is not fast enough.