Meta's MuseSpark 1.2 Gains Embodied Skills: Robot Navigation, Table Cleaning, and Visual Coding
量子位 · wechat · 2026-08-21
Meta has showcased the multimodal and embodied capabilities of MuseSpark 1.2, demonstrating its ability to control robots and generate interfaces from visuals.
Architecture and Demos:
The system features a two-layer architecture. An upper-layer MuseSpark variant handles high-level planning and task decomposition, while a lower-layer VLA (Vision-Language-Action) model executes physical actions. Demos include a robot searching a room for a duck and a dual-arm robot organizing a desk. The model iteratively refines actions based on real-time visual feedback.
Key Capabilities:
- Visual Coding: Can generate websites and games directly from images or videos, ranking first on the DesignArena "video-to-website" leaderboard.
- Embodied Agent: Achieved 54.0% on the ZeroBench visual reasoning test, closely trailing GPT-5.6 Sol (54.6%).
- WildArtifactBench: A new internal benchmark covering 10 complex tasks (e.g., bird call recognition, heart ultrasound analysis, 3D mesh generation) to evaluate cross-modal processing and structured output.
Availability: Currently used internally for media generation and data annotation. Accessible via Meta Model API and MuseCode, with an open-source release confirmed.
More from Embodied
- California startup LUUM built a robot that applies 25 lash extensions per eye — lukas_m_ziegler · 2026-08-21
- Building an offline talking Halloween prop: Stack and challenges — breksyt · 2026-08-21
- Prediction: Cognitive AGI, Autonomous Driving, and Robots Arriving by End of Year — davidpattersonx · 2026-08-21
- VLM Agent for Humanoids Performs Tasks and Conversation — Kangwook_Lee · 2026-08-21
- AdaPT Framework Lets Robots Mimic Federer's Playing Style — Scobleizer · 2026-08-21
- Tienkung Humanoid Runs at High Speed, Corners Like F1 Car — CyberRobooo · 2026-08-21