Safe RL Tooling, Locomotion Pipelines, and the VLM Collapse Problem

News

SPEAR Toolbox Expands with New Environments for Safe RL Benchmarks

Safe RL just got a major quality-of-life upgrade. This 6000-star toolbox (SPEAR) now supports benchmarking across six environments including NuPlan and MetaDrive, making it way easier to compare safety approaches without reinventing the wheel every time.

Latent Diffusion Models for Multi-Sensor Fusion in Robotics: Survey

If you're building camera-sensor combos for drones or embedded robots, this survey maps out the fusion landscape and where Latent Diffusion Models fit in. Good reference for perception teams picking architectures.

Analysis

PEGASUS End-to-End Pipeline for Legged Robot Locomotion

PEGASUS tackles training instabilities and sparse rewards head-on with a modular, automated pipeline that wraps around popular simulators. If you're working on legged locomotion and tired of fighting reward functions, this is worth a close look.

SkillFLANN: Skill-Graph Network for Large-Scale Robot Task Planning

This paper crams 35,000 task demos and 11,000 skills into a training framework with 3D graph relative position encoding. The scale alone makes it a signal that robot pretraining is trending toward internet-scale data curation.

Diagnosing Model Collapse in Vision Language Agent Models

Model collapse is real and this paper puts concrete failure modes on the table for vision-language agents. Builders deploying VLMs in the loop should read this before assuming more data solves everything.

Multi-Modal Subtractive Sensing for Object-Goal Navigation

Object-goal navigation without any pre-built maps, using just egocentric views and active sensing. This kind of mapless approach is exactly what you need for messy real-world deployments where you can't pre-scan every environment.