Tactile Sensing Sharpens, Loco-Manipulation Scales, and Robot Policies Get Safer

Tactile-Visual Fusion Enables Dexterous In-Hand Manipulation Without Ground Truth Labels

This paper fuses tactile and visual signals to handle in-hand manipulation tasks without needing labeled ground truth, which is a massive data bottleneck in dexterous robotics. If this approach generalizes, it cuts a key cost in building manipulation datasets for real hardware.

Whole-Body Loco-Manipulation Framework Achieves Long-Horizon Mobile Tasks in Cluttered Environments

Getting a robot to walk, reach, and manipulate across a sequence of steps in messy real-world spaces is one of the hardest unsolved problems in practical robotics. This framework shows measurable progress on chaining those behaviors without task-specific retraining.

Safe Reinforcement Learning for Robot Manipulation Using Constraint-Aware Policy Optimization

Safety constraints during training have historically killed sample efficiency, but this work shows you can enforce hard constraints without tanking performance. Builders deploying robots in human-adjacent spaces should pay attention here.

Benchmark for Evaluating Generalizable Robot Policies Across Unseen Object Categories

Most robot policy benchmarks let models overfit to a fixed object set, which masks real generalization failures. This benchmark explicitly tests across unseen categories, making it a better proxy for production-environment performance.

Sim-to-Real Transfer for Contact-Rich Assembly via Adaptive Domain Randomization

Assembly tasks with tight tolerances are notoriously hard to sim-to-real transfer because contact dynamics are noisy and hard to model. This adaptive randomization approach narrows the gap without requiring exhaustive manual tuning.

Language-Conditioned Robot Navigation with Hierarchical Subgoal Decomposition

Decomposing a natural language instruction into a hierarchy of spatial subgoals lets a robot navigate longer paths with fewer failures than flat end-to-end approaches. This is a practical architecture pattern worth stealing for any mobile robot product.

Multi-Robot Task Allocation Under Uncertainty Using Graph Neural Networks

GNNs let robots share and reason about task assignments in dynamic environments without centralized coordination, which is exactly what you need for warehouse or construction deployments at scale. The decentralized framing here is what makes this operationally relevant.