Robotics: two humanoids made a bed without talking  May 13th 2026

Watch & Listen First

• Helix 02 Bedroom Tidy (Figure AI, May 8) — two F.03 robots reset a bedroom in under two minutes from a single learned policy.
• Video Friday: Humanoid Robot Production (IEEE Spectrum, May 9) — the 1X Hayward factory tour plus a dexterous-hand demo cracking eggs.
• Practical Applications of Embodied AI and Humanoid Robots (Robot Report Podcast) — Russ Tedrake (TRI) and Scott Kuindersma (Boston Dynamics) on what humanoids can and cannot do today.

Key Takeaways

• Multi-robot collaboration from pixels works. Helix-02 runs two humanoids on one set of weights, each inferring its partner's intent from vision alone.
• Production capacity is the new battleground. 1X opened a 10,000-NEO/year Hayward line, AGIBOT crossed 10,000 units, Atlas fleets pre-allocated to RMAC and DeepMind.
• Binding contracts beat press-release pilots. Humanoid–Schaeffler is a phased 1,000–2,000-unit deployment with a 5-year actuator deal, not a six-week trial.
• Reliability is the rate limiter now, not sim-to-real. The gap between a clip and 99.9% uptime is wider than the hype suggests — every serious operator is saying it out loud.

The Big Story

Figure's Helix-02 runs two humanoids on a shared bedroom task with no message passing · May 8 · Figure AI
→ Two F.03 robots open doors, hang clothes, take out trash, and make a bed together — single VLA policy, ~500 hours of teleop, each robot reading its partner's intent from its own cameras. The technical claim worth taking seriously: no shared planner, no inter-robot comms — coordination emerges from a learned joint action distribution on visual context. Prior bimanual or multi-robot work centralizes planning or hand-codes a protocol; Helix-02 throws both out.

Also This Week

Hyundai's Atlas executes a handstand and an L-sit in its first live production-model demo · May 6 · Bloomberg
→ The whole-body joint control to hold an L-sit for five seconds is the same primitive that keeps Atlas upright through an unplanned shove on a sequencing dolly — stability flex, not circus act; 2026 fleets are already pre-allocated to Hyundai RMAC and Google DeepMind.

1X opens the first vertically integrated US humanoid factory in Hayward, with 10,000 NEO/year capacity · April 30 · The Robot Report
→ Motors, batteries, structures, transmissions and sensors all in-house — the BOM lock-in that will separate survivors from demo-stage outfits when 2027 unit economics get audited.

Humanoid signs a binding 1,000–2,000-unit deployment deal plus a 5-year actuator supply agreement with Schaeffler · May 13 · The Robot Report
→ Phased rollout through 2032 across Schaeffler's plants, Herzogenaurach handling boxes in live production from Dec 2026 — the first 2026 contract that reads like a Tier-1 automotive supply deal, not a robotics pilot.

Japan Airlines starts a two-year humanoid trial at Haneda with Unitree G1 and UBTECH Walker E · April 29 · The Register
→ Baggage, cargo and aircraft towing at one of the world's most regulated airports — at ~$15,400 per Unitree, unit economics roughly an order of magnitude below the US humanoid stack, which will pressure pricing globally.

Boston Dynamics' Spot gets Gemini Robotics-ER 1.6 from Google DeepMind · May · IEEE Spectrum
→ Embodied reasoning up top, classical control underneath — the split-stack (frontier model for planning, learned/MPC for execution) is winning over end-to-end VLAs for any deployment where failures have to be explainable.

AGIBOT ships its 10,000th humanoid and pushes into Japan, Korea and the Middle East · May · The Robot Report
→ 5,000 to 10,000 units in three months — the Chinese stack is scaling volume faster than any Western competitor, and GO-2 pairs that volume with a credible policy layer.

From the Lab

Humanoid Manipulation Interface (HuMI): Whole-Body Manipulation from Robot-Free Demonstrations · arXiv:2602.06643
→ Portable rig captures whole-body human motion (kneeling, tossing, bimanual) and retargets onto a humanoid — 3x faster than teleop, 70% success unseen. The data-pipeline complement to Helix-02: if Figure's policy is the model, HuMI-style rigs are how you get the data without a teleoperator army.

DAIMON-Infinity: an omni-modal robotic dataset with high-resolution tactile sensing · IEEE Spectrum
→ DAIMON's claimed largest omni-modal dataset for physical AI spans laundry to assembly. Vision-only VLAs hit a wall on contact-rich tasks; tactile-aligned data is the modality every serious foundation-model team is chasing.

Worth Reading

• Waymo expands robotaxi coverage to 1,400 sq mi across 11 cities — the only embodied-AI deployment doing a million rides a week is still doing it without a humanoid form factor; hold that in mind when humanoid TAM math gets quoted.
• AGIBOT releases GO-2 foundation model for embodied AI — the Chinese answer to Helix and π0.7; the policy-model field now has three credible camps.

The atoms are catching up to the bits. Two robots made a bed without talking — and a Tier-1 supplier just signed an actuator deal that says they believe it.