Yuanli Lingji DM0.5 tops RoboDojo, open-sources SOTA embodied model
量子位 · wechat · 2026-08-25
Yuanli Lingji's DM0.5 model has topped the RoboDojo embodied intelligence benchmark with a composite score of 24.90 and a 19.34% success rate, leading significantly in memory. Key capabilities include:
- Long Memory & Imitation: Supports 60s memory, learning complex tasks instantly from human demos (like the "Sharingan" copy).
- Fine Manipulation: Achieves millimeter-precision block stacking and flexible object cutting (e.g., cucumbers).
- Robustness: Uses a System 1/2 dual-system architecture to handle viewpoint shifts and interruptions.
Tech Breakdown:
- Data: Fuses 50k hours of real robot ops, 100k hours of ego video, and 1M sqm of simulation data.
- Architecture: Features native long memory, embodied Chain-of-Thought (language-driven reasoning), and alignment-based action supervision.
- Efficiency: Optimized latency from 534ms to 57ms (9.29x speedup).
DM0.5 is fully open-sourced, including weights, framework, and workflows.
More from Embodied
- Hermes Agent Integrated with Pine64 E-ink Display for Offline Voice Interaction — Teknium · 2026-08-25
- Humanoid Robot Breakdown Requires 4 People to Carry It Away — ZehanWang · 2026-08-25
- Unitree and Agibot to produce tens of thousands of humanoid robots this year — chris_j_paxton · 2026-08-25
- Beijing World Humanoid Games highlight a robust robotics ecosystem — chris_j_paxton · 2026-08-25
- Compilation of failures at World Humanoid Robot Games — rohanpaul_ai · 2026-08-25
- Humanoid robot navigates complex 100m obstacle course — CyberRobooo · 2026-08-25