English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Zhiyuan Robot's WALL-B Sets 1,816 Items/Hour in Live Logistics Demo — Embodied AI Starts Counting Cost Per Machine, Output Per Hour

Forum topic · 小凯 · 2026-08-19

Summary

On August 12, 2026, Chinese robotics startup Zhiyuan Robot (自变量) livestreamed a fully autonomous logistics sorting run with no human intervention. A wheeled dual-arm robot with standard grippers, powered by the WALL-B world unified model (WUM), sorted 1,816 parcels per hour at 98% accuracy over one hour on a randomized conveyor line — 45% above Figure AI's Figure 03 benchmark of 1,248 items/hour, which was averaged over 200 hours. The article argues the gap stems from WALL-B's joint multimodal training (vision, audio, language, touch, action) versus Figure's modular Helix VLA stack, and that hardware cost fell over 70% versus Figure's ~$85,000 BOM by dropping five-finger dexterous hands for standard grippers. It also covers Zhiyuan's funding (over RMB 20 billion valuation, backing from Meituan, Alibaba, ByteDance, Xiaomi) and deployments in homes and industrial lines.

On August 12, 2026, Chinese robotics startup Zhiyuan Robot held a fully autonomous logistics sorting livestream with zero human backup. A wheeled robot with dual arms and standard grippers sorted 1,816 parcels in one hour on a conveyor with randomly mixed incoming items, achieving 98% accuracy. For comparison, Figure AI's Figure 03 averaged 1,248 items/hour across 200 accumulated hours — meaning Zhiyuan set the record using less than 1/200th of the time, widening the efficiency gap to 45%. The post frames this not as a muscle-flexing launch event but as a redefinition of embodied AI's value anchor: going forward, "humanoid + dexterous hands" is no longer the default optimum; the cheapest structure that can work stably is.

Three Dimensions of Lead Over Figure

| Dimension | Figure 03 | Zhiyuan's Solution | Gap | |---|---|---|---| | Efficiency | 1,248 items/hour (200-hour average) | 1,816 items/hour (1-hour livestream peak) | +45% | | Hardware cost | ~$85,000 BOM | 70%+ lower than Figure | -70% | | Scene complexity | Highly uniform parcels | Random mix of boxes, foam containers, cylinders, soft bags | Orders of magnitude higher |

The root of the gap is not the number of arms but the brain. Figure 03 uses the Helix VLA architecture, stitching vision, language, and action into separate modules — data loses fidelity across each module boundary. Zhiyuan's WALL-B (released April 2026) uses a World Unified Model (WUM) architecture that trains vision, hearing, language, touch, and action jointly in one network.

Details visible in the livestream illustrate the point:

  • When two parcels overlapped, the robot didn't grab directly; it first judged occlusion, separated the parcels, then handled each
  • Soft bags were nudged, unfolded, flattened, and re-aligned before surface-unit calibration
  • Heavy boxes triggered dual-arm cooperative carrying, pushing from the side to reduce damage risk
  • These on-the-fly judgments are not pre-set trajectories but strategies WALL-B generates in real time after perceiving physical laws.

    Why This Route Wins

    The post argues this is not a "cheaper than Americans" story but one of the ceiling crossing between hardware complexity and software capability.

    On cost: a five-finger dexterous hand has 16 degrees of freedom, each joint needing precision motors, sensors, and control — dexterous hands alone account for 14–18% of a humanoid's BOM and are the first bottleneck to mass production. Logistics sorting only requires "pick up and place" — no walking, no 16-DOF fine manipulation. Two arms with standard grippers, with quick-change grippers ("like F1 tire changes"), can adapt a whole production line in 30 minutes.

    But simpler hardware requires a strong brain — WALL-B's real battlefield:

  • Native multimodality: vision, language, and action trained jointly in one network, eliminating module boundaries
  • Physical-law prediction: predicts causal effects of actions on objects and environments
  • Zero-shot generalization: combines existing physical knowledge into new strategies for unseen boxes, bags, and odd-shaped items
  • A Goldman Sachs survey of 14 Chinese robot companies (May 2026) reached a similar conclusion: most currently prefer wheeled chassis with 2–3 finger grippers, covering 70–90% of industrial applications.

    About Zhiyuan Robot

  • December 2023: founded in Shenzhen
  • 2024: released WALL-A
  • April 2026: released WALL-B, the first WUM-architecture model, with a Series B led by Xiaomi's strategic investment arm
  • April–June 2026: completed B+, B++, and C rounds back-to-back; post-money valuation surpassed RMB 20 billion
  • Investors include Meituan (Series A lead), Alibaba (A+ lead), ByteDance (A++ lead), Xiaomi (B lead) — reportedly the only Chinese embodied-AI company independently led by all four internet giants — plus China Mobile, the National AI Industry Investment Fund, Sequoia China, IDG, Source Code, Fortune Capital, CICC, and 30+ others
  • Founding team: CEO Wang Qian (Tsinghua EE, among the earliest researchers to introduce attention mechanisms into neural networks; PhD in Robotics Learning at USC) and CTO Wang Hao (Peking University computational physics PhD, former head of the Fengshenbang LLM team at IDEA Research Institute).

    Commercialization is progressing on two tracks:

  • Home: 149 RMB robot cleaning service with 58.com's Daojia; an "robots entering homes" program has put WALL-B-powered robots into real households since April
  • Industrial: production lines at a BMW (Brilliance) seat and interior supplier; partnership with SF Express to bring embodied foundation models into industrial logistics
  • Industry Implications: From Concept-Chasing to Deployment

    The post cites a Securities Times commentary: when the embodied-AI industry starts seriously calculating the economics of cost, efficiency, and stability, the signal is clear — the industry no longer blindly pursues "looking human."

    Similar pragmatic shifts are spreading industry-wide:

  • Some companies drop rarely-used pinkies for four-finger dexterous hands to cut cost significantly
  • In relatively flat warehouse and inspection settings, wheeled robots with better stability and endurance are being re-embraced
  • Bipedal forms prone to falling are no longer a fixation
  • Embodied AI competition has crossed the "can it work" demo stage and entered the industrial deep water of "who can do the work cheaper and more efficiently." For Chinese manufacturing this is a structural amplifier: complete supply chains lower hardware costs, while massive real workloads feed back into foundation model iteration. Cheapest hardware + smartest brain is the engineering paradigm to track in Chinese embodied AI over the next 12 months.

    Sources

  • When the embodied robot industry drops its "looking human" obsession — 163.com
  • 1,816 items/hour, 45% efficiency lead over Figure AI — new.qq.com
  • New global record for domestic embodied AI: beating Figure AI at 30% cost — news.qq.com
  • Zhiyuan Robot Technology (Beijing) — Baidu Baike
  • Valuation over RMB 20 billion: four rounds in two months — leaderobot.com
  • Zhiyuan: end-to-end technical deep-dive — finance.sina.com.cn
  • 149 RMB cleaning: robots get schooled by ordinary households — sports.sohu.com

Tags

#embodied-ai#robotics#wall-b#zhiyuan-robot#figure-ai#logistics-automation#world-unified-model#humanoid-robots

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633653