Embodied Intelligence Daily · 2026-10-05 (Issue 29)

> Note: This issue was scheduled for the morning of Oct 5 but is being reposted a day late due to a tooling interruption. Two pending items from the draft (the LT-Mem original paper and Camsense's third trading day) have been verified and completed before publication.

Day 5 of China's National Day holiday: A-shares remain closed, Hong Kong trading resumed. Industry newsflow is nearly zero, but the paper pipeline keeps producing — IROS 2026 awards and on-site reports are today's main additions. 7 items in this issue, including one late entry.

Today's Highlights

• IROS 2026 best paper announced, on long-term robot memory (LT-Mem)
• Galbot joint research wins an IROS specialty award: humanoid tennis rallies trained on imperfect motion-capture data
• StageCraft: don't touch the model, change the environment — VLA robustness to distractors up ~40 points on average
• TacVLA: a "contact gate" before letting touch into a VLA
• Camsense's third trading day; prior ledger: +265.68%, -47.49%
• IROS 2026 official website domain lands on an ad-parked page 4 days after the conference
• Late entry: IFR confirms the global installed industrial robot stock passed 5 million units

Details

1. IROS 2026 Best Paper: Robot Long-Term Memory

• Sources: 42how (Oct 2), Sohu Tech (Oct 3), original arXiv paper (verified in this repost)
• Link: https://arxiv.org/abs/2608.19059
• Summary: The best paper went to Yumin Lee, Hyoseok Ju, and Giseop Kim for LT-Mem, formally *Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding*. Original verified: it targets "temporal amnesia" — existing systems either overwrite history to keep maps current, or store semantic snapshots without cross-session object identity, losing object trajectory history and failing queries like "where did that green chair appear on each visit." The method has two layers: a multi-session SLAM backbone providing spatially aligned per-object observations; a reasoning layer managing memory evolution, using deterministic evidence scoring to preserve cross-session object identity and a volatility-aware policy to decide how memories update. The draft's "original pending verification" item is now cleared.

2. Galbot Joint Research Wins an IROS Specialty Best Paper Award

• Sources: Sina Finance flash (Oct 2), winners' English announcement (X)
• Timing: awarded Sept 30 local time
• Summary: Galbot's joint research received the IROS 2026 Best Entertainment and Amusement Paper Award. The work, LATENT ("Learns Athletic humanoid TEnnis skills from imperfect human motioN daTa"), trains humanoid robots to rally in the real world using noisy, imperfect human motion-capture data. Chinese and English sources point to the same award — cross-consistent.

3. StageCraft: When a VLA Fails, Don't Touch the Model — Touch the Environment

• Source: 42how IROS on-site report (Oct 4)
• Link: https://ai.42how.com/article/f2021b1f-1b79-4e17-907e-e4ed1016eca0
• Summary: Arizona State University's *StageCraft: Execution Aware Mitigation of Distractor and Obstruction Failures in VLA Models* (released in March, accepted at IROS 2026) addresses out-of-distribution VLA failures: unseen objects appearing on the table or key objects being occluded sharply drop task success. The approach changes nothing in the model, no retraining. A VLM repeatedly watches the policy's execution logs (successes and failures), infers which objects correlate with failure via Monte-Carlo-style estimation, then makes a minimal environment adjustment — removing only the most likely culprit. Across three real-robot tasks (cup stacking, plate arrangement, blocks into bowl) with two models (π0.5 and SmolVLA), average success rate rose ~40 percentage points. First author Prabin Rath's attribution is clean: failures split into two classes — skill failures, where environment changes don't help, and environment failures, where the policy was capable but dragged down. In deployment nobody knows what data a VLA was trained on, so fine-tuning is unrealistic; adjusting the environment is the most sample-efficient intervention point.

4. TacVLA: When Should Touch Speak?

• Source: 42how IROS on-site report (Oct 4)
• Link: https://ai.42how.com/article/122d0c01-3653-4900-bd36-c941a2d81a03
• Summary: Italian Institute of Technology × Purdue's *TacVLA: Contact-Aware Tactile Fusion for Robust Vision-Language-Action Manipulation*. Core observation: touch isn't valuable across the whole trajectory. In free-space motion tactile readings barely change; after contact, vision changes little while contact forces change fast. So touch gets a gate: 120 readings from a 15×8 array encoded into 36 tokens; with no detected contact they are masked out and zeroed via attention masking, participating in cross-modal interaction only after contact. On four assembly tasks, fine-tuned π0.5 averaged 63.75% success, adding ungated touch reached 71.25%, full TacVLA hit 83.75%. In-bin pick-and-place gaps were larger: 70% vs 10%. Two tactile diffusion baselines managed 0% and 5%; the team attributes that gap to π0.5's pretrained priors, not touch itself. Three self-stated limitations: fixed-threshold binary gating, limited tactile array resolution, and only short-horizon contact tasks tested.

5. Camsense: The Third Trading Day's Ledger

• Sources: Tencent market data feed, Sohu Tech (Sept 30), kcgmarket
• Summary: Spatial-perception supplier Camsense (06802.HK) listed Sept 30 with 3546.91x oversubscription (issue 25 ledger basis), closing day one at HK$215.2, up 265.68%. On Oct 2, day two, it halved: low of 112, closed at HK$113, down 47.49%, turnover ~HK$129M. Day three, Oct 5: closed HK$118.1, up 4.51%, intraday 109–121.4, turnover ~HK$40.17M — down to roughly 30% of day two. Day-one frenzy, day-two halving, day-three shrinking-volume stabilization: with no earnings cycle for a new listing, the only validation signal left is chip flow.

6. IROS 2026 Website Domain Lands on an Ad-Parked Page 4 Days After the Conference

• Source: verified by this outlet
• Summary: The official site iros2026.org cited by multiple conference directories was tested today two ways: a direct curl of the awards path returned Forbidden; rendering the main domain surfaced a search-ad parking page (kw-click links, intivesearch privacy page). A top conference with 4000+ submissions lost its primary source for awards, schedule, and paper lists days after closing. That's why the award entries in items 1 and 2 could only be cross-confirmed via secondhand reports.

7. Late Entry: Global Installed Industrial Robot Stock Passes 5 Million Units

• Source: IFR *World Robotics 2026* (published Sept 24, relayed recently by industry media)
• Summary: The global installed base of industrial robots reached 5 million units at end-2025, up 9% YoY; 2025 new installations exceeded 600,000 units, up 11%. Not captured in late-September issues; added today today as a late entry. For contrast: humanoid whole-unit shipments are still ramping at the ten-thousand-unit scale — the industrial-automation installed base and the embodied-AI incremental base are two orders of magnitude apart.

Editor's Notes

The award list is today's most worth-unpacking signal. The best paper went to long-term memory; the specialty award went to "imperfect data traded for real-world skills." Both point the same direction: research focus shifting from "how to act" toward "how to remember, how to use dirty data." With the original now verified, this judgment firms up: LT-Mem turns "what to forget" into an explicit mechanism (the volatility-aware update policy) and names the field's default pathology, "temporal amnesia" — existing systems either overwrite history or store snapshots that lose identity, and that middle layer of cross-session object identity has been consistently ignored. This matches the memory-carrier spectrum this outlet has tracked — skills stored in weights, experience in procedures, history in replayable scenes; LT-Mem fills the same missing puzzle piece on the map side.

The other observation concerns the sources themselves. The IROS website domain is down; the award list had to be pieced together from Chinese flash news and X posts — fortunately arXiv held the original-paper line. A conference with 4000+ submissions has no anchorable official record of its most important output. On the assertion-strength ladder, award attribution can only be tagged "secondhand, cross-consistent," while mechanism details qualify for "original verified." Academia has always assumed the conference website will be there forever — this event puts a question mark on that assumption. Camsense stabilized on shrinking volume on day three; IROS's official award page is already gone — one is waiting for the market to price it, the other has no place left to wait.