Embodied AI Daily · 2026-10-07 (Issue 31)

Holiday wrap-up. On the industry side: Agility Robotics held its investor day in New York, entering the final stretch before SPAC pricing. On the Chinese consumer side: 100 Lingxi X2 robots deployed into retail stores and Hubei adding embodied robots to subsidy programs. The research highlight: the arXiv 10-05 batch contained four companion infrastructure papers for World Action Models — attack, acceleration, verification, and audit — a full set.

Key Points

• Agility Robotics Investor Day (Oct 6, New York, livestreamed): Digit 5, expansion strategy, Schaeffler and Foxconn on stage. The ~$2.5B SPAC merger's pricing window is imminent.
• Four WAM infrastructure papers (arXiv 10-05 batch): an adversarial patch drops FastWAM success to 0%, one-step distillation yields RealtimeWAM, FAVOR fills the gap of "what to recover after execution drift," and an audit paper finds instructions are encoded but do not control actions.
• Zhiyuan deployed 100 Lingxi X2 units into 100 Aishida stores; Hubei added embodied robots to trade-in subsidies starting Oct 1.
• RV-ICL built on RPent: LIBERO-PRO 92.6% → 96.5%. The September open-source infrastructure is attracting third-party follow-ups.
• H-JEPA (LeCun among authors): hierarchical world models lift Visual AntMaze from 18% to 73%.

Details

1. Agility Robotics Investor Day: A ledger for the "first pure-play humanoid stock"

• Sources: Agility Robotics official announcements / Yahoo Finance
• Link: https://www.agilityrobotics.com
• Time: 2026-10-06 20:30 ET (livestream)

Agility held an Analyst & Investor Day in New York with three agenda centers: Digit 5, capacity expansion strategy, and customer/partner perspectives from Schaeffler, Foxconn, and AUVSI. The company announced its merger with Churchill Capital XI in June at a valuation of roughly $2.5B, expected to raise over $620M for expansion, becoming the first US pure-play humanoid robot listing. Order book: over $300M in multi-year orders as of May. Per investor-day messaging (company social media newsflash), Digit has moved over 125,000 totes at customer sites — versus the official 100,000 milestone in November 2025, a ~25% gain in five months. In late September the company also disclosed it is exploring wheeled and other form factors, no longer betting solely on bipedal designs. Schaeffler, a shareholder since November 2024, attended with gearbox collaboration and additional investment (details pending official disclosure).

One-line takeaway: for the first time, the US market is handing "what is a humanoid robot worth" to public pricing. Agility's argument is not demos — it's operating numbers.

2. Four WAM infrastructure papers: a new term grows a full support ecosystem

(Four papers appeared in the same arXiv 10-05 batch; read together as one signal.)

TAPDreamer (arXiv 2610.06814): Adversarial attacks on world-action models. Using only a public vision encoder, an attacker can construct a fixed local patch without querying the victim model. The patch covers ~6.5% of the input yet drops FastWAM's success rate from 97.7% to 0.0% across 40 LIBERO tasks, and from 90.8% to 0.0% on 50 RoboTwin tasks. Same-area random patches only reduce success to 81.5% and 79.2%. Mechanistically, patch-induced attention–value interactions broadcast an almost uniform representation shift beyond the patch, stable across tasks.

RealtimeWAM (arXiv 2610.06617): Inference efficiency. Teacher-Anchored Consistency Distillation (TACD) enables one-step action generation, addressing the gap that "small local consistency error doesn't guarantee final action accuracy." Cross-Expert Wavefront Pipelining (CEWP) overlaps video and action experts, sharing KV cache per chunk with a single sync before action-attention consumption — removing both the multi-step denoising and serial expert-waiting bottlenecks.

FAVOR (arXiv 2610.06280): Execution verification and online recovery. Once execution deviates from a WAM's predicted future, remaining actions are void, and replanning from the drifted state rarely recovers. FAVOR's clean observation: the answer is already in hand — the WAM's predicted future frames are the states it intended to pass through. Keep those frames as anchors; a verifier compares observations against anchors to flag deviation, then a VLM translates anchors into a short corrective instruction. LIBERO 97.85% → 98.10%, LIBERO-Plus 72.60% → 72.98%. The numbers are modest; the value is that "what to recover after execution drift" finally has a systematic answer.

Encoded but Not in Control (arXiv 2610.06235): Grounding audit. With scenes unchanged, swap instructions — to another visible object, arbitrary nouns, or irrelevant sentences. Tested VLAs and most WAMs still go for the scene-associated original target. Linear probes recover the modified instruction target from intermediate representations (language is encoded), while attention analysis shows instruction tokens contribute weakly to action generation. The conclusion is in the title: encoded, but not in control.

3. Zhiyuan's 100 Lingxi X2 robots into 100 stores: consumer-facing guides

• Sources: China Securities Journal (via Sina Finance) / Information Times
• Link: https://finance.sina.com.cn/jjxw/2026-10-03/doc-initxwvx3848836.shtml
• Time: 2026-10-03

Ahead of the National Day holiday, Zhiyuan sent 100 Lingxi X2 units into 100 Aishida stores nationwide as shopping guides — greeting, explaining, matching needs, directing customers, coordinating with human staff on closing sales. From Sept 28, joint three-day, three-city store livestreams began. Meanwhile Shoucheng Holdings' "Taozhu Maker Lab" opened 8 retail stores plus 4 pop-ups; the Life Hub (Heshenghui) store drew ~10,000 weekly visitors, with holiday daily traffic expected above 2,000. One of the best-selling embodied products in-store: a 49,000 RMB stair-climbing mobile robot, nearly 10 units per month. On the policy side: Hubei added embodied robots, cooking robots, exoskeletons, and companion robots to trade-in subsidies from Oct 1; robot dogs, AI glasses, and exoskeletons already carry a 15% national subsidy since mid-August.

4. Unitree GD01 crewed mecha debuts in Tianjin: marketing department's holiday

• Source: Jinhai Binhai (via Toutiao)
• Link: https://www.toutiao.com/article/7691252643875897910
• Time: Posted 2026-09-30, exhibit Oct 1–7

A nearly 3-meter-tall red crewed transforming mecha, GD01, made its public debut at Tianjin TEDA Life Mall. Unitree calls it "the world's first mass-producible crewed transforming heavy mecha" — "first" and "mass-producible" are vendor claims without third-party verification, treated as weak assertions. Semi-open crewed cockpit; on-site tie-ins with consumer devices from UBTech, DeepBlue, and others at 5% off, with rental and purchase options. The same day, a Yangtze Delta embodied-AI 6S experience center opened in Shangyu, Shaoxing (Oct 1): eight zones, 30+ products, with Zhiyuan, UBTech, Fourier, and others exhibiting; holiday bookings were hot. Both samples point to the same thing: the holiday consumer face of embodied AI — robots becoming mall "shows" and shelf goods.

5. RV-ICL: first ecosystem follow-up on RPent infrastructure

• Source: arXiv 2610.06843
• Link: https://arxiv.org/abs/2610.06843
• Time: 2026-10-05

Instead of stuffing demonstration videos into prompts as long sequences, RV-ICL turns them into a hierarchy the agent can navigate. The hierarchy is built from sub-events of the demonstration (grasp, release, etc.), refined from key frames down to phases, moments, and clips, exposed via read-only tools. The agent reads coarse levels before planning and re-enters the hierarchy during execution, loading only the segment for the current subgoal. Training-free; one demonstration per task suffices. Built directly on RPent (a Tsinghua × Infinigence × Zhengxing open-source embodied agent infrastructure, covered in the Sept 22 issue): LIBERO-PRO 92.6% → 96.5%, LIBERO-Plus 86.7% → 95.8%.

A clear stance on memory format: video as an experience carrier, keeping structure beats converting to text. Textual memory records "what was done"; a video hierarchy preserves "how it was done" — contact details determine whether a grasp succeeds, and key frames are exactly where those details get lost.

6. H-JEPA: LeCun's hierarchical world model

• Source: arXiv 2610.06805
• Link: https://arxiv.org/abs/2610.06805
• Time: 2026-10-05

End-to-end training of "a hierarchy of action-conditioned JEPAs." Each layer predicts further into the future in its own latent space; the top level optimizes progress toward goals, its predictions becoming subgoals for the next level's planner. When a dataset's compositional factors evolve on separated timescales, higher layers naturally discard fast, unpredictable details and retain slow, task-relevant state. On Visual AntMaze, a three-level hierarchy lifts success from 18% to 73% with less planning compute. Extended to DROID real-robot videos with inverse-dynamics supervision, offline planning fidelity also improves. Yann LeCun is among the authors.

This is the first time the JEPA line has made "hierarchical" an end-to-end trainable recipe rather than a dashed line in an architecture diagram.

7. InterMimicGen: humanoid manipulation data flywheel

• Source: arXiv 2610.06850
• Link: https://arxiv.org/abs/2610.06850
• Time: 2026-10-05

A three-step closed loop: retarget motion-captured human-object interaction datasets into robot references, preserving whole-body coordination and hand-object relations; train a generalist tracking policy in physics simulation, larger and more diverse than prior humanoid trackers; then open the flywheel — each round makes "task-semantic-preserving small edits" to where interactions happen and how bodies execute, keeping only variants that complete the task in simulation as seeds for the next round. Over rounds, small edits compound into wide coverage of sparse original demonstrations without degrading task semantics or motion quality. An extension of the MimicGen lineage to humanoid whole-body manipulation. Another point on the data-collection spectrum: not collecting more data, but letting data edit itself.

8. ArtifactArena: Elo ranking by what models build

• Source: arXiv 2610.06511
• Link: https://arxiv.org/abs/2610.06511
• Time: 2026-10-05

An open platform where frontier models take on physically grounded hardware-software co-design challenges: build robots that actually battle in an arena. Three harnesses iterate model outputs via text descriptions, physics simulation feedback, and match data, ranking frontier models by head-to-head tournament Elo. Framework and tournament infrastructure are open-sourced for continuous community submissions — a "non-saturating living testbed." The evaluation target shifts from "what the model says" to "what usable things the model builds."

(Two more worth noting: VLA-ZO [2610.06271] uses zeroth-order optimization for deployment-time adaptation, lifting task success under camera viewpoint drift from 48.27% to 63.58% with 25× less adaptation time, no backpropagation anywhere; DexForge [2610.06331] uses differentiable physics to turn human video demonstrations into high-fidelity trajectories for seven dexterous hands, improving success by 35–53 percentage points over baselines — another physics-grade entry in the lossy-interface business of retargeting.)

Editor's Observations

1. WAM's word-week has grown into a category-week. The Sept 20 paper week noted "WAM becomes a word" (12 papers that week); the Oct 6 issue covered three papers from the Oct 2 batch on the fuel problem (using videos without action labels). Today's 10-05 batch adds four — all infrastructure: someone attacks it (TAPDreamer, 97.7% to 0), someone compresses it (RealtimeWAM, one-step generation), someone verifies it (FAVOR, anchor-based recovery), someone audits it (Encoded: instructions live in representations but not in control). The sign of a concept maturing isn't rising citations — it's attackers, compressors, verifiers, and auditors showing up simultaneously. The most striking of the four is the audit: swap the instruction, and the robot still grabs the original scene object; language is encoded but doesn't control action. Together with TAPDreamer it points to the same thing — success rates measure execution, not understanding. Benchmarks are assertions; deployment is execution. The embodied side just collected two more samples of that thread. A falsifiable prediction: WAM adversarial robustness will appear as the first embodied entry in some safety benchmark within 12 months.

2. Two holiday "deployments" side by side, both using the cheapest verification bandwidth. In the US, Agility put its 125,000-tote operating number directly on the investor-day table — with SPAC pricing imminent, deployment figures are valuation arguments. In China, nearly a thousand robots performed in malls, 100 Lingxi units entered cookware stores, Hubei wrote embodied robots into trade-in subsidies. Neither chose the strictest verification (controlled experiments, long-horizon audits); both chose the cheapest (foot traffic, transactions, subscription multiples). No contradiction: mass-deployment verification has two phases — cheap signals filter what deserves strict verification, then strict verification endorses what deserves mass production. The holiday ends tomorrow night, and two bills are pending: the official pricing of Agility's merger, and the FLAG set in yesterday's issue — the failure rate and hourly-wage bill for Golden Week robot "shows." Time to settle up.

(Sources for this issue: 13 effective web searches (1 rate-limited, 1 empty) + 30 papers pulled via arXiv API + abstract-by-abstract verification of key papers; Agility investor-day details follow official PR and livestream announcements, social-media numbers are tier-annotated.)