MaxxBench introduces 89 DeepSWE tasks with minor instruction or code changes. Models clustered around 74% on DeepSWE show a 23 pp spread.
Key points
89 DeepSWE tasks with small changes to instructions or starting code.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Sources
DefFlow – goal-conditioned 3D world action modelling announced
DefFlow is a goal-conditioned 3D World Action Modelling approach that learns to manipulate deformable objects into diverse goal shapes using 3D point flows. It targets ropes, granular objects, and cloth for precise 3D shape control.
Key points
DefFlow: goal-conditioned 3D World Action Modelling approach
Learns to manipulate deformable objects into diverse goal shapes using 3D point flows
Sources
Agensi – Skill Marketplace for AI Agents Launches with 18,000+ Skills
Agensi is a skill marketplace that lets AI agents load expert-built capabilities on demand, adding over 18,000 new skills instantly. The platform offers free and paid options and integrates with agents via a single connector line. Engineers can purchase or sell skills, with a 70/30 revenue split and a flat 30% platform fee.
Key points
18,000+ new skills available instantly for AI agents.
Revenue split is 70% to creators, with a flat 30% platform fee on each sale.
Sources
- Original post
- Agensi: The Skill Marketplace for AI Agents - Linked page
CLARA clarification framework for robotics announced
CLARA clarification framework for robotics announced, a cognitively-driven system that asks users clarifying questions to ground ambiguous requests. It refines vague directives into structured instructions and generates visual renderings with a diffusion pipeline.
Key points
Name: Conversational cLArification for Robotic Arrangement (CLARA)
Work conducted in early 2025
Sources
RAX scaling online exploration at CoRL2026 poster 119
RAX is an agentic robotics system that learns by scaling online exploration. It will be presented as a Late-Breaking Result poster at CoRL2026 on Nov 9, 12–1:30 PM CT, poster #119 in Griffin Hall, JW Marriott Austin.
Key points
Presentation: Late-Breaking Result “Scaling Online Exploration for Robot Agents” at CoRL2026, Mon Nov 9, 12–1:30 PM CT, poster #119.
System: RAX is an agentic robotics system that learns by scaling online exploration.
Sources
ORA Open Robot-Use Agents presented at CoRL2026
ORA: Open Robot-Use Agents will be presented at the CoRL2026 Late-Breaking Results session on Nov 9, 12–1:30 PM CT, Poster #130. It runs on DeepSeek 4.1-Flash and can zero-shot control robots from visual input.
Key points
Presentation: CoRL2026 Late-Breaking Results, Mon Nov 9, 12–1:30 PM CT, Poster #130, Griffin Hall.
Model: powered by DeepSeek 4.1-Flash, zero-shot robot control from visual input, no VLA.
Sources
- Original post
- github.com - Linked in the post
- haoailab.com - Linked in the post
