Latent Skill Stacks, Synthetic HRI Pipelines, and the Sim-to-Real Transfer Frontier

Latent Skill Composition for Generalizable Motion Planning Across Morphologies

This paper introduces motion planning as a problem of learning to compose primitive skills in latent space, which is important because it offers a way to generalize locomotion across different robot morphologies without retraining from scratch. Builders working on legged robot stacks should pay attention to the modularity here.

Ultra-HumanVid: Synthetic Data Generation for Human-Robot Interaction Videos

Ultra-HumanVid tackles the lack of large-scale human-robot interaction video data by generating synthetic training sequences with physically plausible contact. This is a bottleneck that has been holding back manipulation policy learning, and synthetic data pipelines like this could unlock faster iteration.

Hierarchical Task Decomposition for Sim-to-Real Manipulation Transfer

This work proposes a hierarchical approach for sim-to-real transfer in robotic manipulation, decomposing tasks into easier sub-problems that generalize better when deployed on hardware. The gap between simulation and the real world remains the hardest part of building deployable robots, and this paper offers a practical framework for narrowing it.

HeCoFuse: Heterogeneous Collaborative Perception via Cross-Sensor Fusion

HeCoFuse addresses multi-robot coordination by fusing heterogeneous sensor modalities across a team of robots for collaborative perception. If you're building multi-agent systems for warehouses or search-and-rescue, efficient sensor fusion is a high-leverage problem worth studying.

Learning Robot Manipulation from Unpaired Human Demonstration Videos

This paper looks at learning manipulation policies directly from human demonstration videos without requiring paired robot trajectories, which dramatically lowers the data collection cost. For builders scaling imitation learning, this kind of approach could cut the annotation bottleneck significantly.

RobotWin: LLM-Based Zero-Shot Robotic Task Planning and Evaluation

RobotWin explores using large language models for zero-shot task planning in robotic manipulation, testing how well LLMs can reason about spatial relationships and tool use without task-specific training. The results are mixed but the approach is worth tracking as foundation models get better at grounded reasoning.