Edge-Efficient Policies: Shrinking Models and Growing the Stack

Research

LLM-Generated Skill Specifications for Autonomous Manipulation

Rather than learning every manipulation policy from scratch, this paper shows how LLMs can generate robot code on the fly based on object context. This could be the transformer for rapid prototyping, catching edge cases that brittle behavior trees miss.

Efficient Deployment of Vision-Language-Action Models via Structured Pruning

Vision-Language-Action models are strong, but usually too expensive for the robot's onboard compute. This paper proposes a way to compress these massive policies to run locally without killing reliability, which is critical for making these models usable off the grid.

Placing Functional Affordances at the Center of Embodied Visual Representations

Instead of generic object recognition, they're training models to see the world as a collection of functional interaction points. This connects perception directly to motor output, cutting out the middle compute that usually slows down closed-loop control.

News

ROSCon 2025: ROS 2 Jazzy Jalisco Release and Desktop Improvements

ROS 2 Jazzy Jalisco is officially out, featuring significant tooling upgrades and performance improvements. Anyone running a production ROS stack needs to look at the migration path now.

Tools

NVIDIA Isaac Sim: Automating Contact Rich Interactions and Implicit Actuation

Building digital twins is slow if you have to micromanage physics contacts. NVIDIA just dropped quality-of-life updates for simulating implicit actuation and contact handling in Isaac Sim, which means less debugging time when closing the sim-to-real gap.