Affordance Models Get Smarter, Multi-Robot Coordination Scales, and Dexterous Hands Learn From Less

Language-Conditioned Affordance Prediction for Open-World Robot Manipulation

This paper ties language instructions directly to affordance maps, letting robots figure out where and how to interact with objects they haven't seen before. That's a meaningful step toward robots that generalize instead of just memorizing task-specific policies.

Scalable Multi-Robot Task Allocation via Decentralized Graph Neural Networks

Decentralized GNN-based coordination lets robot fleets allocate tasks without a central bottleneck, which is exactly what warehouse and logistics deployments actually need. The scalability results here are worth paying attention to if you're building multi-agent systems.

Learning Dexterous In-Hand Manipulation from Sparse Human Demonstrations

Getting useful dexterous manipulation policies from just a handful of human demos is a big deal for anyone who can't afford to collect thousands of teleoperation hours. This approach cuts data requirements significantly without tanking policy quality.

SafeNav: Constraint-Aware Navigation for Mobile Robots in Dynamic Environments

Real-world robot navigation keeps failing because static safety constraints don't hold up when environments change. This framework bakes constraint awareness into the planner in real time, which matters a lot for robots operating near people.

One-Shot Imitation Learning for Tool-Use Tasks via Cross-Embodiment Transfer

Cross-embodiment transfer for tool use means you can train on one robot morphology and deploy on another, which dramatically lowers the cost of building new manipulation capabilities. One-shot performance on novel tools is the headline result here.