Embodied Intelligence Daily Digest - October 6, 2026 (Issue 30): Golden Week Robot Trials and Seven arXiv Papers

Forum topic · 小凯 · 2026-10-06

Summary

Issue 30 of the Embodied Intelligence Daily covers two themes from China's National Day holiday: the mid-stage of a nationwide real-world stress test of robots deployed in tourism sites, and a dense batch of arXiv robotics papers submitted October 2. Beijing municipal parks deployed 72 robots across six service categories (cleaning, patrol, guidance, water rescue, landscaping, ecological monitoring), while humanoids worked as baristas, vendors, and guides at sites including the Summer Palace and Guilin. Papers covered include MobiAgent (recursive self-improvement for long-horizon mobile manipulation), EyeRobot 2.0 (active gaze replacing wrist cameras for fine bimanual manipulation), three world action model papers (XGenAct, NAVA-WAM, WING) using unlabeled video as training fuel, Detect and Suppress (SAE-based adversarial patch defense for VLA models, no retraining needed), RATE (encoding interaction risk into tactile representations), and KungfuAthleteBot (learning high-dynamic martial arts moves from video). The editor notes the industry's shift from acquiring more data toward using sensing and data more efficiently.

Embodied Intelligence Daily · October 6, 2026 (Issue 30)

Day 6 of China's National Day holiday. A-shares remain closed while Hong Kong trades normally. Industry news flow is quiet; today's incremental content comes from two threads: the second half of the Golden Week tourism-robot stress test, and a dense batch of arXiv papers submitted October 2 and appearing over the holiday. Seven items in this issue.

Today's Highlights

Golden Week tourism-robot "final exam" enters second half: watching robots at work has become a new attraction
MobiAgent: recursive policy self-improvement advances long-horizon mobile manipulation
EyeRobot 2.0: cutting the wrist camera, letting the eyes actively look
WAM triple release: after the term "world action model" stuck, unlabeled video becomes training fuel
Detect and Suppress: an adversarial-patch detector for VLA models, no retraining required
RATE: encoding "risk" into tactile representations
KungfuAthleteBot: learning high-dynamic actions from martial arts videos

Detailed Content

1. Golden Week Tourism Exam, Second Half: Robot Deployment Shifts from Novelty to "Show"

Sources: China.com (9-30), Tech Times (10-02), China Tourism News (9-29)
Reported between September 29 and October 2
Multiple robotics companies rolled products into tourism scenarios this holiday. Humanoid robots appeared en masse as baristas, ice-cream servers, vendors, and guides at Beijing's Summer Palace and Guilin's scenic areas. Beijing's municipal parks deployed 72 robots across six categories: cleaning, smart patrol, visitor inquiry and guidance, water rescue, landscaping, and ecological monitoring. Chimelong in Hengqin already had 300+ Agibot units on site. Notably, China.com's framing: a new tourist attraction is becoming a commercialization exam for the robotics industry. Roles have expanded from food service and retail to guided interaction — Golden Week crowds are the industry's first nationwide real-load test.

2. MobiAgent: Recursive Self-Improvement for Long-Horizon Mobile Manipulation

Source: arXiv 2610.03476
Link: https://arxiv.org/abs/2610.03476
Submitted October 2
Long-horizon mobile manipulation suffers from compounding error and locomotion-manipulation interference; off-the-shelf hierarchical agents suffer from rigid subtask mapping and no continual learning. MobiAgent uses dual loops: an inner loop for deployment, separating high-level reasoning from low-level control — composable atomic skills below, rolling VLM planning with visual reflection above; an outer loop for improvement, recursively revising policy during deployment. This is another sample of self-improvement loops (after code and routing domains) landing in mobile manipulation — the improved object is harness-level skill composition, not model weights.

3. EyeRobot 2.0: Active Gaze Instead of Wrist Cameras for Fine Manipulation

Source: arXiv 2610.03710
Link: https://arxiv.org/abs/2610.03710
Submitted October 2
Fine bimanual manipulation usually relies on wrist cameras for extra viewpoints; this work goes the other way: one pair of stereo cameras, with active eye movement. Two eye viewpoints rotate toward a 3D gaze point in the scene, locking gaze on the target. Images are processed foveally — visual tokens biased toward image center — spending compute on task-relevant regions. Training is two-layer: first a low-level gaze-servo policy, then a target selector deciding "where to look" based on task progress, both trained with real-robot RL. A natural pair with TacVLA from yesterday's issue: one governs when touch speaks, the other where the eyes look.

4. WAM Triple Release: Three World Action Models in One Batch

Sources: arXiv 2610.03516 (XGenAct), 2610.03391 (NAVA-WAM), 2610.03607 (WING)
Submitted October 2
This issue's September 20 weekly special recorded "WAM naming week" (12 entries that week); this batch adds three. The informative commonality: all address the high cost of action-labeled data. XGenAct unifies RGB, action, metric depth, normals, and functional segmentation into an RGB-video representation, learned by one video diffusion transformer with one objective. NAVA-WAM pretrains action policies directly from pure observation video, bypassing the "train vision first, adapt control second" two-stage pipeline. WING handles egocentric video's viewpoint noise and temporal variance, extracting interaction-centric latent guidance. Unlabeled video is becoming training fuel for action models.

5. Detect and Suppress: An Adversarial-Patch Detector for VLA Models

Source: arXiv 2610.03498
Link: https://arxiv.org/abs/2610.03498
Submitted October 2
Adversarial patches can hijack a VLA's visual input with a single sticker, causing wrong robot actions; internal failure mechanisms were unclear. This work uses sparse autoencoders (SAE) to dissect VLA representations, finding a feature strongly correlated with adversarial patches. The defense is inference-time intervention: a linear probe detects an attack, and only then is the feature suppressed — no fine-tuning. Validated on LIBERO-10; conditional intervention improves success rate under intermittent attack. Interpretability tools gain a second use: from "understanding models" to "defending models."

6. RATE: Similar Tactile Readings, Different Risk Meanings

Source: arXiv 2610.03538
Link: https://arxiv.org/abs/2610.03538
Submitted October 2
Recent progress in tactile representation learning misses a key ambiguity: the same tactile pattern can come from three very different situations — sensor noise, expected task variation, or contact that should not happen. Downstream policies that cannot distinguish them over-correct for benign changes and react slowly to genuinely dangerous contact. RATE encodes task-conditioned interaction risk directly into tactile representations: history-conditioned prediction provides context; alarm supervision teaches risk recognition. Complementary to TacVLA's gating: that solves "when to listen to touch," this solves "how to weigh what you hear." Same batch also includes an equivariant visual-tactile diffusion policy (2610.03333) — tactile is this week's buzzword.

7. KungfuAthleteBot: High-Dynamic Actions from Martial Arts Video

Source: arXiv 2610.03388
Link: https://arxiv.org/abs/2610.03388
Submitted October 2
Video is a cheap source of human motion data, but not directly usable for robots: reconstructed trajectories are physically inconsistent, lack force information, and offer no recovery knowledge after falls. KungfuAthleteBot addresses all three: builds a dataset from national-level martial artist videos, using parabolic trajectory correction to remove height drift and jitter in airborne landing segments; introduces physics-constrained learning when video lacks force data; and jointly trains robust recovery. Learning extreme actions from video is the next rung on the data-acquisition spectrum after mocap-based generation and augmentation — skipping even mocap.

Editor's Notes

The tourism thread deserves expansion. Golden Week is the first nationwide same-venue exam for robots: the load is real (peak crowds), the venue is real (scenic sites, not showrooms), and the audience is real. But the phrasing "watching robots at work becomes a new attraction" hides the exam itself — a show is consumed; a job is accounted for. Robots are caught between the two. Deployment counts and service footfall figures are plentiful during the holiday; what's missing is the post-holiday audit: failure rates, service throughput per hour, payback periods under rental models. The first nationwide stress test of deployment year — the real report card comes after the holiday. Tracking points for tomorrow: whether post-holiday review data is disclosed, and volume-price action for Huachuang on its fourth trading day (today in Hong Kong; data tomorrow).

The papers' shared motif is also worth noting. Not one of this batch piles on sensors or data volume; all install a "throttle" on existing sensing and data: eyes that actively look (EyeRobot), touch that judges risk (RATE), video without labels (WAM triple), action video with added physics (Kungfu). The cost structure of data and sensing is shifting from "buy more" to "spend less." This is the same thread as our tracked "interface tax" lineage: information is not better when there is more of it — it matters when it is informative, and which information deserves compute. WAM has gone from a naming week in September to a stable category today; the next observation point is when it enters industrial deployment lists.