LLM Reward Shaping, Latent Actions from Video, and Tactile Legs on Rough Ground

Research

Predictive Traffic Model for Safe Unmanned Vehicle Intersection Navigation

Title: Predictive Traffic Model Boosts Unmanned Vehicle Intersection Navigation

This paper presents a model that predicts traffic participant trajectories to help unmanned vehicles navigate intersections safely. It's relevant because urban navigation remains one of the hardest unsolved problems where perception meets planning at high stakes.

3DGS vs NeRF: Representation Choices for Robot Manipulation

Title: Efficient 3D Gaussian Splatting vs NeRFs for Robot Manipulation

Compares 3DGS vs NeRF representations for robotic manipulation tasks. Data representation choices in the perception stack have outsized impact on real-world performance, so this kind of systematic benchmarking helps builders pick the right foundation.

LLMs as Reward Designers for Reinforcement Learning Locomotion

Title: Using LLMs as Reward Designers for RL-based Locomotion

Leverages language models to design reward functions for locomotion policies. This could dramatically reduce the manual engineering effort needed to tune legged robot behaviors and accelerate sim-to-real transfer cycles.

Spatial-Temporal Attention for Long-Horizon Robot Manipulation

Title: Simultaneous Spatial-Temporal Attention for Long-Horizon Manipulation

Addresses long-horizon manipulation with a new attention mechanism that captures both spatial and temporal dependencies. Long-horizon tasks are where most manipulation systems fall apart, so any architectural improvement here is worth watching.

Latent Action Pretraining from Unlabeled Video for Robot Learning

Title: Latent Action Pretraining from Unlabeled Video for Robot Control

Pretrains latent action representations from videos without action labels. This taps into the massive pool of unlabeled video data, which could be a game-changer for scaling robot learning beyond expensive teleoperated demonstrations.

ViTaL: Vision-Tactile Locomotion for Quadrupeds on Rough Terrain

Title: ViTaL: Tactile-Informed Locomotion for Quadrupeds on Rough Terrain

Combines vision and tactile sensing for quadruped navigation on deformable and rough surfaces. Terrain-aware locomotion that doesn't rely purely on vision is critical for real deployment where ground truth is messy and contact-rich.

Events

ROSCon 2025: Call for Proposals and Registration Open

Title: ROSCon 2025 Announced with Call for Proposals

ROSCon is the premier community event for ROS developers and roboticists. Getting on stage or attending is still one of the highest-signal ways to understand what's actually shipping in production robotics stacks.