Embodied AI Daily · October 1, 2026 (Issue 26)

First day of China's National Day holiday: Hong Kong markets closed, Chinese industry news in a quiet window. International developments didn't take a break — Microsoft's VLA foundation model, Figure's third-party evaluation, and IDC's half-year report all landed in the two days before the holiday. This issue focuses on the international front.

Today's Highlights

• Microsoft open-sources Rho: a 5B dual-arm VLA foundation built around "adaptation is the bottleneck"
• IDC half-year report: global humanoid robot H1 shipments near 25,000 units, +432% YoY, China at 77.9%
• Figure releases RIDGE evaluation: Astra and Opus 5.5 near-equal on average, but trade wins task-by-task
• arXiv submission wave: "harness" root appears in a robotics paper title for the first time (RoboHarn-Evo)
• DeepRobotics advances IPO; two quadruped market reports released same day disagree from second place onward

Details

1. Microsoft Rho: An open-source VLA foundation that turns "adaptation" into a product

• Source: Microsoft Research
• Links: https://arxiv.org/abs/2609.38164 · https://microsoft.github.io/rhobotics/
• Timeline: arXiv submitted 09-29, project page live, front-page HN exposure today

A 5B-parameter open-source VLA family focused on dual-arm manipulation. One Rho-base pretrained foundation, midtrained into three hardware variants: I²RT YAM Box, UR AI Trainer (2×UR5e), and Franka FR3 Duo. The training recipe is four stages plus an optional fifth: Phi-Phy vision-language base → attach a 542M flow-matching action expert (3,200+ hours of robot data, 48 B200 GPUs, 90% robot data) → per-hardware midtraining (~300 hours each) → task fine-tuning → post-deployment online learning.

Key numbers: midtraining halves task fine-tuning data requirements; on the BusyBox benchmark Rho-YAM-Box hits 90% success, on par with π0.5, while GR00T N1.7 gets only 53%; a single LIBERO policy handles 40 tasks at 97.9% average. Most interesting is online adaptation: FlowDAgger uses 15 human corrections to pull FR3 Duo tube assembly from 30% to 70% — the foundation weights stay frozen throughout, training only a lightweight latent policy; correction actions are inverted into regression targets, so online learning only shifts "where generation starts," while actions still come from the untouched offline model. The same holds in Meta-World simulation: an undertrained policy plus 50 scripted corrections per task lifts six hard tasks from 57.5% to 73.3% on average.

Weights, code, data, and both benchmarks are fully open-sourced on HuggingFace at microsoft/rho.

2. IDC half-year report: H1 shipments near 25,000 units, China at 77.9%

• Source: IDC (reported by PChome, Blue Whale Finance, et al.)
• Link: https://technode.com/2026/09/29/china-accounts-for-77-9-of-global-humanoid-robot-shipments-in-h1-idc-says/
• Timeline: published 09-29

Global humanoid robot H1 shipments approach 25,000 units, +432.1% YoY; market size exceeds US$740 million, +322.7% YoY. China shipped over 19,000 units, +426.3% YoY, 77.9% of the global total. At the vendor level the report names only one company: AGIBOT shipped over 8,600 units, roughly 45% of China's total by the 19,000-unit figure. IDC judges the industry to be "moving from technical validation toward commercial application," with scenarios expanding to industry, retail, and home.

Reference point: full-year 2025 figures were Unitree at 5,500 and AGIBOT at 5,168 (institutional figures via smzdm). AGIBOT's half-year 8,600 units already exceeds its full-year 2025 total; the market jumped from thousand-unit to ten-thousand-unit scale within two years. Against institutional forecasts for 2030, there are still two orders of magnitude to go — all the delivery pressure sits in the second half.

3. Figure RIDGE: what averages hide, task-level data exposes

• Source: Figure AI Research Blog
• Link: https://www.fig.inc/blog/astra-opus-5-5-and-other-frontier-models-demonstrate-jagged-performance-across-sota-agentic-tasks-from-web-browsing-to-robotics/
• Timeline: 09-29

Figure evaluated seven models — GPT-6 Astra, Claude Opus 5.5, GPT-5.6, Fable 5, Kimi K3, and others — task by task across five domains: web (VisualWebArena), driving (Bench2Drive-VL), industrial workflows (IndEgo), assembly (Assembly101), tabletop manipulation (VLABench).

The verdict is called "jagged performance." On VisualWebArena, Astra scores highest at 139/177 with Fable 5 close behind at 138/177 — but in every model pair, the lower-scoring model solves tasks the higher-scoring one cannot: Opus 5.5 scored 124/177 yet solved 25 tasks that Astra didn't. Model version updates reshape the task profile without moving the average. Same in physical domains: the model with higher domain average is worse on some task categories. The accompanying RIDGE dataset is open-sourced with full item-level results and model outputs across five domains.

One line: benchmark averages are marketing figures; task-level variance is the deployment figure.

4. arXiv submission wave: "harness" enters titles, evaluation audits enter VLA

• Source: arXiv cs.RO new submissions
• Links: https://arxiv.org/abs/2609.37583 · https://arxiv.org/abs/2609.37771 · https://arxiv.org/abs/2609.38172
• Timeline: submitted 09-29

Three threads worth watching.

RoboHarn-Evo (2609.37583): "harness" appears in a robotics paper title for the first time. A dual-loop harness evolves "hierarchical physical knowledge" from physical experience — task knowledge handles subtask scheduling and completion criteria; action knowledge handles object-relative geometric policies and their physical effects. Foundation model weights untouched; knowledge grows outside the model.

Faster and Better? (2609.37771): an audit of the evaluations of VLA acceleration methods themselves. Finds that training-free acceleration methods measure higher success rates than baselines on benchmarks; tracing anomalous-gain tasks to root causes splits into implementation bugs and design limitations. First systematic excavation of the acceleration-evaluation pit.

Counterfactual Video Generation (2609.38172): counterfactual video generation for humanoid loco-manipulation data augmentation, connecting to the world-model synthetic data line.

5. DeepRobotics advances IPO; two quadruped reports clash same-day

• Source: Phoenix Tech, Sina Finance
• Timeline: reported 09-30 (quadruped dual reports published 09-21)

DeepRobotics, one of Hangzhou's "six little dragons," enters the IPO queue: tutoring started 2025-12 and completed in five months; a 366-page prospectus filed with the Shanghai Stock Exchange. Here's the metrics story — Frost & Sullivan's metric for the prospectus claims "No. 1 globally in 2025 quadruped robot industry-application revenue," while IDC's 1H26 hardware revenue ranking puts Unitree first. And on 09-21 itself, Counterpoint and IDC released quadruped market reports the same day: both agree Unitree is first, but rankings from second place onward diverge entirely.

Shipment units, hardware revenue, industry-application revenue — three rulers, three outcomes.

6. Roboflow: GPT-6 Astra is the strongest vision model tested

• Source: Roboflow Blog
• Link: https://blog.roboflow.com/gpt-6-astra-vision/
• Timeline: originally published 09-18, revived on HN this week

GPT-6 Astra scores 82.1% mAP@50 on Roboflow Vision Evals, leading Qwen3.8 Max by 5.4 points and GPT-5.6 Sol by 13.7 points; no competitor closes the gap even at high inference tiers. Testing covered box prompting, segmentation, re-identification, and robot control. Its launch-time computer-use capability spills over into general vision tasks.

Editor's Notes

Microsoft Rho and GPT-6 Astra delivering in the same week is a face-off between two routes. Astra bets on zero-shot generalization backed by scale; Rho writes "zero-shot still isn't reliable in new environments" into its first paragraph and bets on the middle layer — per-hardware midtraining halves fine-tuning data, and 15 post-deployment corrections lift 30% to 70%. Open-source 5B versus closed frontier models, competing for the same scenarios: on LIBERO Rho-base scores 97.9%, while Astra posts 92.63% on LIBERO-PRO (different benchmarks, not directly comparable, but both scorecards are on the wall). Figure's jagged evaluation cuts from the middle: leading averages don't mean task-level leadership; model selection requires item-level visibility. IROS closes today in Pittsburgh; last week's generalist-vs-specialist debate still has no Chinese-side coverage — the answer isn't in the conference hall, it's in the comparison of these two scorecards.

The metrics story appeared for the third time this week. IDC's 77.9% is clearly defined (shipment units), yet the same agency's 09-21 quadruped report and Counterpoint's same-day report diverge from second place onward; the Frost & Sullivan metric in DeepRobotics's prospectus (industry-application revenue first) is again a different thing from IDC's hardware-revenue metric. Research agencies have gone from one ruler to three; the reader's discipline is unchanged: ask about the denominator first. One more thing worth recording: on the first day of the holiday, the Chinese industry side went silent while the international side dropped Rho, RIDGE, and an arXiv wave within 48 hours — capital-market holidays and technology-side cadence have fully decoupled. Hong Kong stays closed tomorrow, so the Chinese side likely stays quiet; internationally, watch the IROS closing digest and community reproduction feedback on Rho.