Language-Guided Control Tightens, Deformable Object Manipulation Advances, and Locomotion Policies Get More Robust

Language-Conditioned Robot Policies With Improved Semantic Grounding

Connecting natural language instructions to low-level robot actions remains a hard alignment problem. This work pushes semantic grounding further, which matters for anyone building general-purpose manipulation systems that need to follow varied human commands without retraining.

Learning to Manipulate Deformable Objects via Topological State Representations

Deformable objects like cloth, ropes, and cables are notoriously hard for robot policies because state is ambiguous and high-dimensional. Using topological representations to track state cuts through that complexity and gets closer to practical deployment in real environments.

Robust Legged Locomotion Under Terrain Uncertainty With Adaptive Policies

Locomotion policies that break on unexpected terrain are useless outside a lab. This paper targets robustness under uncertainty directly, which is the gap keeping most quadruped and biped systems from real-world deployment.

Scalable Imitation Learning for Long-Horizon Robot Tasks Using Hierarchical Decomposition

Long-horizon tasks are where most imitation learning approaches fall apart as compounding errors kill performance. Hierarchical decomposition is a practical structural fix that makes these pipelines actually scale without needing massive datasets.

Contact-Implicit Trajectory Optimization for Dexterous In-Hand Manipulation

Most dexterous manipulation planners treat contact as a discrete event, which creates brittleness. Contact-implicit formulations handle this more gracefully and are a key step toward hands that can regrasp and reorient objects reliably.

Sim-to-Real Transfer for Soft Robotic Grippers Using Domain Randomization and Learned Residual Models

Soft grippers have huge practical upside for unstructured environments but sim-to-real gaps are worse than for rigid robots. Pairing domain randomization with learned residual corrections is a concrete recipe that actually closes the gap for compliant hardware.

Multi-Modal Sensor Fusion for Real-Time Obstacle Avoidance in Dynamic Environments

Single-modality perception fails in cluttered, fast-changing spaces which is most real-world settings. Fusing vision, depth, and proprioception in real time and keeping latency low enough for reactive control is still an open engineering challenge this paper directly addresses.