【直撃取材】ヒューマノイドは現場で使えるか／フィジカルAIの実装は物流でどこまで進んでいるのか／人型の実装がまだ少ない理由／Mujin、Geek+、川崎重工業の最新展示／中国勢の強み

Summary

Physical AI in logistics (goal & rationale)

• The video visits an International Logistics Center to evaluate how far physical AI has progressed—i.e., AI that controls robots operating in the real world.
• A central theme is why robots struggle in logistics more than in factories:
  • Factories handle repeatable, mass-produced items.
  • Logistics is highly variable: many SKUs/product types arrive daily, which creates complex picking/sorting demands and constantly changing operational conditions.

Robotics industry “common OS” / platform

• One exhibit features an industrial robotics platform described as a “common OS” for robotics.
• A robot arm uses:
  • A vision system (camera mounted near the top/orange camera area).
  • Hand-mounted sensors to recognize what it sees.
• The system maintains awareness even after items are placed, tracking:
  • Location and counts (e.g., how many boxes are in each position).

Why humanoids are rare in logistics—and what’s different about this approach

• A robotics researcher notes it’s rare in Japan to see a humanoid demonstrated at a logistics-focused event using real logistics products.
• The video explains practical hurdles:
  • Logistics tasks are complex and require recovery from failure (i.e., “retrying”).
  • Adding humanoid-level complexity can slow operations, though the demo claims speed can be maintained through testing.

Geek+ (CEO Kato): humanoid for picking/sorting in warehouses

• Geek+ presents a new logistics/factory humanoid, framed as an evolution from their AGV “Roomba-like” systems (with large deployments).
• Demo workflow and behavior:
  • Uses working stations to dispense containers and to provide product selection data to the humanoid.
  • The humanoid performs picking and sorting, placing items into bins/baskets assigned to different people/orders.
• End-effector (gripper) emphasis:
  • Designed for versatile grasping, not a single-purpose tool.
  • Uses tactile/pressure sensors on fingertips plus a camera on the hand to estimate grip force/position and reduce dropping.
• Learning and data-driven improvement:
  • The system repeatedly attempts tasks; when it fails (e.g., drops), it can accumulate data and improve over time.
• Design choices vs classic bipedal humanoids:
  • Includes wheels at the bottom rather than a typical two-speed bipedal design.
  • This aims to reduce battery consumption for 24/7 operation.

Labor shortage as the business driver

• The video argues humanoid/automation value is not only about outpacing human speed, but also about insufficient staffing:
  • Examples suggest many temporary workers do not show up, forcing automation to meet shipment deadlines.
• Even if humanoids are slower per individual motion, continuous operation can yield major throughput gains.

Proof of Concept (POC) model

• A commercialization tactic highlighted is proactive POCs:
  • Companies test how the workplace must change to use robots effectively—not just whether the business case works financially.
• Geek+ references large-scale field results:
  • Hundreds of thousands to millions of picks daily globally.
  • POCs in China with real customers, including deployments of around ~100 units.

Mujin (and other systems): robustness and warehouse architecture

• Another demo describes a pallet-moving “4-shuttle” system:
  • Unlike conventional stacker-crane approaches (where one stop can halt everything), this design adds fault tolerance.
  • It uses multiple rail-running square AGVs (RGVs) under pallets to diversify shutdown risk.
• The demo also mentions earthquake-resistant structure design, reflecting Japan-specific reliability requirements.
• It compares pallet-container management to more manual forklift-style workflows, noting driver availability is declining.

Kawasaki Heavy Industries: humanoid-adjacent logistics and AI grasping

• Kawasaki shows a warehouse solution involving:
  • Double depalletizing
  • Transporting goods via delivery trucks to stores
• AI usage is central:
  • Vision AI recognizes boxes/SKUs and produces instructions for picking.
  • Picking has many variations because contents and grasp points differ.
• Bottleneck: training data and annotation
  • Traditional methods rely on human annotation to label grasp points on images, which is slow and time-consuming.
• “Model learning / lawless learning” style approach
  • Kawasaki claims a faster method for generating training data:
    • About 23 seconds to teach “pinching” for one sample.
    • An AI data point can be created every ~3 seconds, enabling roughly 100 repetitions quickly.
  • The approach emphasizes capturing real fingertip contact data, since simulating friction/holding uncertainty is difficult.
• Retry behavior in practice
  • The robot can handle different object softness (e.g., plastics vs. toothpaste) by restarting/trying again when it fails.
• Strategy: humanoid form factor vs practicality
  • Kawasaki notes a focus on humanoid/social robots in general, but for logistics it prioritizes systems that are automatable and commercially deployable.
  • A key point: fingertip dexterity is especially hard because human fingers are both highly capable and costly to replicate, helping explain why full humanoids aren’t widespread yet.

China vs Japan strengths (users vs robotics capability)

• The researcher contrasts:
  • Japan’s strength: industrial robotics and adapting the environment to robots (“robot-friendly environment”).
  • China’s advantage: broader user adoption (how many robots are deployed/used), including humanoid-related use cases—even if some industrial technology is closer in capability across regions.
• Overall takeaway:
  • Physical AI can shift the balance once robots learn enough to operate effectively even in environments not originally designed for robotics.

Open challenges discussed

• Humanoid/dexterous robots still struggle with:
  • Grasping complex shapes
  • Handling heavy items
  • Dealing with variable products
• The hardest part is described as not sorting or even gripping, but reliably getting items into the box (final placement reliability).
• Hardware details under discussion:
  • Differences in camera setup (e.g., a reference to “3 eyes” vs potentially 2 or 4) tied to realism and capability.

---

Key speakers / sources

• Tetsuya Ogata — robotics/AI professor and logistics robotics expert (main researcher perspective)
• Mr. Kato (Geek+) — CEO of Geek+
• Kawasaki Heavy Industries (person in charge / technical head context mentioned) — logistics warehouse demonstration and AI grasping approach (referenced as “Yoshikawa” and “Mr. Kawasaki” in segments)