Embodied AI Daily Digest · 2026-10-03

> Issue 27 · Day three of the National Day holiday (Saturday). Industry sources have gone quiet, and the Hong Kong market was closed October 1, with a half-day session on October 2 to digest the Camsense crash. But arXiv doesn't take holidays — a batch of cs.RO papers submitted just before the October 1 deadline (US Eastern) appeared at 8 AM Beijing time yesterday. This issue is paper-driven, with one industry anchor.

Today's Highlights

• WAM (World-Action Model) fires three times in one day: SkeleWAM, UniWAM, and World Motion Models appeared simultaneously along three representation routes — with WMM earning a NeurIPS 2026 Spotlight.
• InterEvolve: humanoid robots solve new tasks via test-time evolution of reward programs — no weight retraining.
• RPG: an embodied agent's self-improvement targets the skill library and system prompt, not model weights.
• HumanoidToolBench: GR00T N1.7 exposes the gap between "picking the right tool" and "completing the task" across 18 tool-use tasks.
• Camsense (Huanchuang) crashed 47.49% on its second trading day — a two-day roller coaster in pricing the physical-AI narrative.

Detailed Coverage

1. World Motion Models: Compressing a Dynamic World into SE(3) Trajectories

• Source: arXiv 2610.01742 (UC Berkeley; Qianqian Wang / Trevor Darrell / Angjoo Kanazawa)
• Link: https://arxiv.org/abs/2610.01742
• Timeline: submitted 2026-10-01, NeurIPS 2026 Spotlight
• Summary: The paper proposes sparse SE(3) pose trajectories as the minimal modeling primitive for 4D worlds — articulated bodies, humans, hand-object interactions, camera motion, and robot states/actions all unified in one shared space. Training uses per-token noise-level flow matching. Its sharpest claim is a reduction statement: future prediction, action completion, MPC, inverse kinematics, cross-embodiment retargeting, and policy learning are all "different masks on the same network." Where SkeleWAM argues video representations retain control-irrelevant appearance, WMM goes further and stops predicting pixels entirely.

2. SkeleWAM: Robot Joints + Object-Centric + Interaction Points — a Skeleton as the World Model

• Source: arXiv 2610.02120 (Juyi Sheng / Hua Wang / Mengyuan Liu)
• Link: https://arxiv.org/abs/2610.02120
• Timeline: 2026-10-01
• Summary: Existing WAMs predict video or visual latents, keeping geometric relations only implicitly. SkeleWAM constructs a sparse 3D skeleton online from RGB-D and proprioception as a unified geometric state for action generation and future prediction; predicted future skeletons add geometric supervision to action learning without visual reconstruction. At inference it maps the current skeleton plus a language instruction directly to actions, with Medoid Action Consensus as a corrective. The route in essence: retreat the world model from pixels to geometry.

3. UniWAM: Complementary Flaws of VLA and WAM in One Architecture

• Source: arXiv 2610.02054
• Link: https://arxiv.org/abs/2610.02054
• Timeline: 2026-10-01
• Summary: VLAs have semantic understanding but lack grounding in world dynamics; WAMs have spatiotemporal priors but fail to reason under distribution shift. UniWAM puts a physical reasoner, world generator, and action predictor into a unified architecture, mixing human egocentric and robot data via a data-cleaning pipeline. Two design points stand out: low-level actions are represented in natural language (semanticized action interface), and the pretraining recipe assigns VQA, egocentric, and robot demonstrations to appropriate components. Post-training uses future-visual noise augmentation to reduce dependence on precise future prediction.

4. InterEvolve: Evolving Reward Programs at Test Time, No Retraining

• Source: arXiv 2610.02196 (Zhuo Lin / Sirui Xu / Yu-Xiong Wang / Liang-Yan Gui et al., in the FB body-prior line of work)
• Link: https://arxiv.org/abs/2610.02196
• Timeline: 2026-10-01
• Summary: The core hypothesis: a broad base controller already holds most capabilities needed for new tasks — it only lacks a planning-control interface expressive enough and executable enough. InterEvolve builds this interface with two pieces: an object-aware forward-backward behavior foundation model that compiles new rewards into loco-manipulation behaviors on a frozen body prior, and tasks written as staged reward programs with completion conditions and tunable constants. An LLM agent revises program structure in context, a numerical optimizer tunes constants, and each candidate is verified in parallel simulation before entering a "verified program skill library" — evolution is confined to the reward-program layer; everything outside is frozen.

5. RPG (Reconstruct, Practice, Go Real): Fixing the Skill Library and System Prompt, Weights Untouched

• Source: arXiv 2610.02204 (Yen-Jen Wang / Weirui Ye / Rocky Duan et al.)
• Link: https://arxiv.org/abs/2610.02204
• Timeline: 2026-10-01
• Summary: RPG identifies manipulation capabilities from offline data, constructs practice tasks in simulation, and diagnoses failures using execution feedback, privileged simulator state, and dataset videos — producing three things: new reusable symbolic skills, refined old skills, and a revised system prompt. All changes pass cross-task evaluation; both individual candidates and merged revisions must be tested before retention. At test time, a multimodal LLM goes on-robot with the final system prompt and skill library. No step in the pipeline updates gradients — the object of self-improvement is the harness layer of the execution system.

6. HumanoidToolBench: What Happens When GR00T N1.7 Picks Up an Unfamiliar Tool

• Source: arXiv 2610.02089 (Seoul National University, snu-pi)
• Link: https://arxiv.org/abs/2610.02089
• Timeline: 2026-10-01
• Summary: 18 tasks, three scenes, three execution levels, two tool-set modes, plus 3.1k demonstration episodes (ToolBook; split between simulation and real Unitree G1). Seven policies run in simulation, three on hardware. The result is a fault line: a huge gap between being able to select tools and being able to complete tasks. Probes on GR00T N1.7 yield two findings — tool-selection accuracy drops noticeably on unseen tools, and the robot continues its original task when given irrelevant instructions. The latter is both a robustness flaw and an honest sample of instruction following.

7. Camsense (Huanchuang): +265.68% on Day One, −47.49% on Day Two

• Source: JRJ / Guandian / Eastmoney (financial press reports)
• Link: https://stock.jrj.com.cn (coverage)
• Timeline: reported evening of 2026-10-02
• Summary: The full ledger — IPO priced at HK$58.85, raising HK$682M with BYD as a cornerstone investor; gray-market gain of +190.23%; on 9-30 the stock opened at HK$187.6 and closed at HK$215.2 (+265.68%), market cap HK$20.78B, turnover near HK$300M; on 10-02 it crashed 47.49%, nearly halving market cap — roughly HK$10B evaporated in two days. Camsense's actual business is lidar modules for robot vacuum cleaners (Camsense visual positioning sensors), having just turned profitable last year. Hong Kong was closed for the National Day holiday on 10-01; the half-day on 10-02 was enough to digest the sentiment — the market gave the "physical AI first stock" label exactly one day of credibility.

Briefs

• FlashDexRetarget (arXiv 2610.01849): RL-accelerated dexterous-hand retargeting with separate left/right actor-critic networks and FlashSAC off-policy algorithm; a 50-motion benchmark that lowers the cost of converting human demonstrations into robot data — a supply-side answer to retargeting's lossy-interface problem.
• 3DROID (arXiv 2610.01744): a renderable 3D Gaussian dataset with per-scene reliability measures; a calibration-aware pipeline anchors reconstructed scenes back into the robot's metric workspace.
• ChunkVLA-AM (arXiv 2610.01856): OpenVLA-OFT deployed on a FAIRINO FR3 additive-manufacturing station with eight-step open-loop action chunks and inter-chunk closed-loop control — a practical case study in the economics of action chunking for factory VLAs.

Editor's Notes

The September 20 issue counted "WAM" appearing 12 times in one week of cs.RO papers and declared it a word. This week it entered its next phase: once a term becomes a word, internal routes begin to split. Three papers in one day — SkeleWAM with sparse 3D skeletons, UniWAM unifying architectures, WMM compressing the entire dynamic world into SE(3) trajectory sets. The common move across all three is a retreat — from predicting pixels to predicting geometry. SkeleWAM says it plainly: video representations retain appearance information irrelevant to control. This newsletter has seen this retreat pattern before: visual laziness (L70-U25-G5) was the collapse of task weighting, the draft-paper three-solutions was the collapse of intermediate states, and WMM is the collapse of the representation primitive itself. Hidden in WMM's claim that "six task classes all reduce to different masks on one network" is its most radical assertion: cross-embodiment retargeting is just another mask.

Two other same-day papers answer an older question: at which layer does the R in self-improvement actually occur. InterEvolve's answer is the reward program; RPG's answer is the symbolic skill library plus system prompt. Neither touches gradients anywhere. InterEvolve's "verified program skill library" and RPG's "cross-task evaluation before retention" both build verification gates into the skill-accumulation loop. This newsletter has called this position the harness layer, or the structured-form shape of experience carriers — this week it appeared twice more, as two independent implementations in the same batch of papers. InterEvolve comes from UIUC, and RPG's author list includes Weirui Ye — two entirely unrelated lines converging on the same design is more noteworthy than any single paper's claim.

Camsense's two-day roller coaster adds a third pricing posture to the "base-layer capitalization" ledger this newsletter has been keeping since late September. Benmu's 208x oversubscription with a first-day gain of +6% was restraint under Chapter 18C constraints; Unitree spent its first post-IPO month slowly paying a HK$247.5B validation tax; Camsense's +265% then −47% was the market collecting from day one. Three samples side by side: first-day gain itself becomes a readable signal — it marks the gap between narrative and substance. Camsense's base-layer identity is real (BYD as cornerstone, lidar modules genuinely on robots), but between the "physical AI first stock" hat and the "robot vacuum supply chain" body, the market completed its repricing in one day. Until the holiday ends, these three samples will stay in the same frame — and the first verification point worth tracking afterward may be this: whether institutions step in after the halving carries more information than how high the first day went.

(End)