How Six Robots Built a 3.5‑Meter Great Wall in 15 Hours Using VLA+World Model

Less than an hour before the opening of the 2026 World Artificial Intelligence Conference (WAIC), six robots performed a continuous 15‑hour operation to assemble a Great Wall replica composed of more than 80,000 bricks, measuring 3.5 m long, 1.5 m wide and 1.1 m high. Four desktop robots handled fine‑grained part assembly while two wheeled Apex robots transported and positioned the assembled components.

DM0.5: Sub‑millimeter Precision and Broad Generalization

DM0.5 enables robots to act within a 0.1–1 mm range, surpassing human physiological jitter (≈0.3–1 mm). The precision derives from high‑accuracy perception, precise actuation, and autonomous execution, delivering steady, accurate, micro motions. DM0.5 also generalizes across varied components and autonomously recovers from errors.

Three structural upgrades differentiate DM0.5 from its predecessor:

Historical information fusion : introduces up to one‑minute memory of key frames while retaining fallback to current‑frame decisions when history is unavailable.

Embodied reasoning task expansion : adds 11 self‑regressive tasks that jointly supervise instruction understanding, temporal reasoning, and action generation, improving long‑range instruction compliance and motion continuity.

Dynamic trajectory alignment : uses dynamic programming to monotonically match predicted actions to real trajectories, learning task regularities rather than data‑collection cadence, resulting in smoother and more robust core motions such as grasping and placing.

Inference runs on an Action‑Chunk basis; a default flow‑matching of 10 steps produces a 50‑step action block, reaching 10 Hz on a single RTX 4090 and 20 Hz on an H100.

Benchmark Performance of DM0.5

DM0.5 leads multiple leaderboards:

RoboChallenge Table30 V2 real‑robot test: 43 % overall success rate, composite score 54.42 (SOTA).

LIBERO simulation benchmark: average score 99.0, surpassing π0.5 and GR00T N1.7.

RoboTwin2.0 dual‑arm benchmark: 93.5 points.

Navigation benchmarks R2R and RxR (Val‑Unseen): top success rate and SPL.

DW0.5 and the VLA + World‑Model Post‑Training Loop

To address assembly failures caused by brick‑craft precision, the team employed the DFOL2.0 post‑training framework together with the DW0.5 world model. DW0.5 learns a "learned environment" for VLA: it samples candidate actions from current observations, generates future visual rollouts, and feeds them to a Value Expert that scores each trajectory’s success probability and value. The scores become dense reinforcement‑learning feedback, enabling the RL coach (CFG‑RL) to update DM0.5 weights. Only a small amount of real‑robot rollout is needed for continual calibration.

DW0.5 makes three key design choices:

Action as strong prior : uses a group‑diagonal attention mask in a Mixture‑of‑Tokens architecture to bind each video frame to its corresponding action group, cutting off cross‑talk between unrelated frames and actions.

Simulating both success and failure : its data pool mixes public embodied datasets, self‑collected data, internet videos, first‑person human activity recordings, and real‑robot rollouts, explicitly including both successful and failed trajectories, as well as RoboChallenge real‑world success/failure data.

Value Expert provides dense feedback : transforms sparse task rewards into per‑step success‑probability scores, enabling effective RL updates.

These designs yield top scores on evaluation suites:

EWMBench: 4.66 points, first place.

WorldArena: 73.54 points, first place.

RoboTwin2.0: 93.3 points (SOTA).

Synergy of DM0.5 + DW0.5: A Stronger Application Model

In the combined system, DM0.5 serves as the base policy model, DW0.5 predicts action consequences and supplies value feedback, and reinforcement learning links the two. The 15‑hour, 80,000‑part Great Wall build acts as a stress test: any memory loss or motion drift would accumulate error, yet the DM0.5 + DW0.5 pipeline delivered uninterrupted, high‑precision assembly.

Source code and model checkpoints are publicly available:

DM0.5 GitHub: https://github.com/dexmal/opendm

DM0.5 Hugging Face: https://huggingface.co/Dexmal/DM05

DW0.5 GitHub: https://github.com/dexmal/opendw

DW0.5 Hugging Face: https://huggingface.co/Dexmal/DW05-Base