WAIC Live: 6 Robots Autonomously Assemble 80,000-Block Great Wall in 15 Hours, Embodied Foundation Model DM0.5 as Core

At the 2026 World Artificial Intelligence Conference (WAIC), Yuanli Lingji, in collaboration with StepFun, launched a world-first autonomous robot assembly challenge: 6 heterogeneous robots (4 desktop robots + 2 humanoid wheeled robots Apex) worked continuously for 15 hours, using over 80,000 micro building blocks to autonomously construct a Great Wall model measuring 3.5m long, 1.5m wide, and 1.1m at its highest point. The entire process was free of manual remote control or preset scripts, driven entirely by Yuanli Lingji's newly released embodied general-purpose foundation model DM0.5.

Challenge Details and Difficulty

• Precision Requirements: Block interlocking required precise alignment within 0.1–1 mm, while natural human hand tremor is about 0.3–1 mm. Robots needed to operate stably at precision near the physiological limits of humans.
• Collaboration Mode: The 6 robots operated without a central scheduler, dynamically dividing tasks through communication negotiation, sharing progress in real time, and coordinating spatial occupancy, achieving decentralized multi-agent collaboration.
• Long-Horizon Stability: Continuous operation for 15 hours demanded stable output from the robots to avoid cumulative errors.

Core Models: DM0.5 and DW0.5

• DM0.5: Yuanli Lingji's self-developed embodied general-purpose foundation model, adopting a VLA architecture with a 4B-parameter multimodal backbone + 680M action expert. Training data reached 150,000 hours (including 50,000 hours of high-precision real-robot operation data and 100,000 hours of first-person scene videos). Compared to the previous generation DM0, data volume increased by 400% and parameter count doubled.
  • Key architectural improvements:
    • Context Abstraction Layer: Provides up to 60 seconds of task process memory, preventing "forgetting step by step" in long-horizon tasks.
    • Embodied CoT: Introduces 11 autoregressive reasoning tasks, expanding training supervision from single action prediction to joint supervision of instruction understanding, temporal reasoning, and action generation.
    • Trajectory Alignment Layer: Uses dynamic programming to monotonically align predicted actions with real trajectories, eliminating individual rhythm differences of collectors and enabling the model to learn task patterns.
• DW0.5: World model, serving as a "Learned Environment" in the post-training loop, used to simulate action consequences and generate success/failure trajectories. Combined with the reinforcement learning coach CFG-RL for policy optimization, it significantly reduces real-robot trial-and-error costs.
  • Key designs: Actions as strong priors (frame-level binding), simultaneous simulation of success and failure, and Value Expert providing dense feedback.

Evaluation Results

• RoboChallenge Table30 V2: DM0.5 achieved SOTA with 43% overall success rate and a comprehensive score of 54.42.
• LIBERO Simulation Benchmark: Average score 99.0, surpassing baselines such as π0.5 and GR00T N1.7.
• Dual-Arm Collaboration Benchmark RoboTwin2.0: Leading with a score of 93.5.
• Navigation Benchmarks R2R/RxR: DM0.5-Nav ranked first in success rate, SPL, and other metrics.
• World Model Evaluation: DW0.5 ranked first on EWMBench with a score of 4.66 and first on WorldArena with a score of 73.54.

Impact and Significance

• Technical Validation: The challenge validated breakthroughs in embodied foundation models for sub-millimeter precision manipulation, long-horizon stable operation, multi-robot collaboration, and open-world generalization, offering the potential for robots to take over approximately 20% of manual, non-standard processes in manufacturing.
• Industry Trend: The "brain + body" combination of Yuanli Lingji and StepFun reflects the industry trend of deep binding between large model companies and robotics hardware companies.
• Open Source: DM0.5 and DW0.5 have been open-sourced (GitHub: dexmal/opendm, Hugging Face: Dexmal/DM05), driving the development of the embodied intelligence community.