Embodied AI Daily · 2026-09-16

Today's Highlights

• Infinigence, with Tsinghua and Shanghai Jiao Tong University, open-sourced APXInf, an edge-side inference engine for embodied AI: π0.5 VLA runs on Jetson Thor at 41ms FP8 latency / 24.3Hz
• Unitree released the fully upgraded G1+ (95,000 RMB incl. tax); the same day, Lei Jun visited Unitree and sparred with a robot
• Agility Robotics launched Digit 5: human-robot collaboration safety as a first-class feature, payload up to ~23kg, 90-min battery with 9-min fast charge
• Xinhua survey: over 70 embodied AI training grounds are operational nationwide; "robot schools" become new-type infrastructure
• AutoNavi with Nanjing, Tsinghua, and Peking Universities released WorldRoamBench: 1,000+ long-horizon samples benchmarking 12 interactive world models

Details

1. Infinigence open-sources APXInf edge inference engine: π0.5 on Jetson with near-lossless FP8 quantization

• Sources: Sina Tech / GitHub (infinigence/ApxInf)
• Link: https://github.com/infinigence/ApxInf
• Date: 2026-09-15

Infinigence, together with Tsinghua University and Shanghai Jiao Tong University, officially open-sourced APXInf, an edge-side inference engine for embodied AI, declaring "embodied AI is highest priority." It optimizes VLA/WAM model inference for Jetson Thor/Orin-class edge devices, implemented in Rust with zero external dependencies. The first release is highly optimized for Physical Intelligence's π0.5 VLA and supports BF16/FP8/INT8 precision. Performance: on Jetson AGX Thor, FP8 latency is 41.16ms (24.3Hz), reaching 26.32ms (38.0Hz) with single-step action generation pruning; RTX 4090 INT8 hits 38.5Hz. Accuracy holds up: on LIBERO-10, FP8 quantized success rate is 92.2% (BF16: 92.8%) vs the π0.5 reference of 92.4% — essentially no regression. An OpenPI-compatible websocket server allows an unmodified openpi-client to connect directly. A companion robot-side repo (RLinf/APXinf-robo) covers Franka + LIBERO evaluation, and an agentic workflow porting new models with an agent — the README describes it as an engine "born of the agentic coding era," with CUDA kernels agentically optimized.

2. Unitree G1+: 2 new neck DoFs, +110% shoulder/waist motor torque, 95,000 RMB incl. tax; Lei Jun visits the same day

• Sources: Sina Tech / iFeng / MyDrivers (ITHome)
• Date: 2026-09-14

Unitree launched the fully upgraded G1+ on September 14, with six upgrades across motion, perception, and intelligence: 2 new neck DoFs for flexible head rotation; a fully upgraded motor system — shoulder/waist peak torque +110% with 72% less heat at equal torque, arm motor torque +43% with 30% less heat, thigh motor heat down 28% at equal torque; and a dual-eye + wide-angle binocular vision system. The standard version is priced at 95,000 RMB including tax; pricing for the research-oriented G1+ EDU was not announced. It retains the 9,000mAh quick-swap battery (~2h nominal), adds a DC54.6V/5.7A external power port for continuous operation, and supports OTA / WiFi6 / Bluetooth 5.2. That evening, Xiaomi's Lei Jun visited Unitree, watched martial arts and brush calligraphy demos, and sparred with a robot — which was pushed back several steps but regained balance without falling. Per 21st Century Business Herald, Shunwei Capital, founded by Lei Jun, was an early Unitree investor. On the capital side: Unitree's market cap once hit 444.9 billion RMB in its first listed month (1,100 RMB intraday on Aug 19), but fell below 490 RMB on Sept 11 to under 200 billion — roughly 250 billion evaporated in a month.

3. Agility Digit 5: safety for human collaboration as a first-class product feature, run-charge ratio from 2:1 to 10:1

• Sources: Agility Robotics official announcement / PR Newswire / Forbes
• Link: https://www.agilityrobotics.com
• Date: 2026-09-15

Agility Robotics launched the fifth-generation Digit, positioned as "the first humanoid robot engineered for safe, scaled human collaboration." The safety architecture is this generation's core: proprietary AI algorithms plus new sensor categories continuously monitor nearby people and autonomously take action — proactive avoidance, stopping motion, or switching to a seated posture; the official demo shows it maintaining balance while carrying load when a rug is pulled from underfoot. Per-carry capacity rises from ~16kg to ~23kg (50 lbs); the new battery provides 90 minutes of runtime with 9-minute fast charging, lifting the run-to-charge ratio from Digit 4's 2:1 to 10:1, which the company says supports 20+ hours of daily operation. Forbes previously reported Digit v5 ships in December. A telling contrast: in the same week, Unitree is competing on motor torque and a 95,000 RMB price, while Agility is competing on collaboration safety and charging economics.

4. Xinhua field survey: 70+ embodied AI training grounds built; "robot schools" become new-type infrastructure

• Sources: Xinhua / CAICT "Embodied AI Training Ground Research Report (2026)"
• Link: https://www.xinhuanet.com
• Date: 2026-09-15/16

Xinhua reports that "robot school"-style embodied AI training grounds are accelerating nationwide. The CAICT report shows: as of the end of June, over 70 training grounds are operational, with 40+ more under construction or planned; coverage spans more than half of provincial-level regions, forming three clusters (Yangtze Delta, Beijing-Tianjin-Hebei, Pearl Delta) and extending into tier 3–5 cities. The supply chain shows a clear inverted triangle: active upstream technology supply vs downstream applications still to be developed, with heavy capital intensity. Two related developments in Hangzhou: on the 16th, the National AI Application Pilot Base (Embodied AI) was inaugurated; Sept 16–17 saw the 2026 Global Embodied AI Developer Conference and the seventh Embodied AI Training Camp open at Xidian University's Hangzhou Institute, attended by Xiong Youjun, CEO of the Beijing Humanoid Robot Innovation Center. Per Leidian Caijing, tens of billions of RMB flowed into embodied AI in H1, with an industry matchmaking event held in Hangzhou on Sept 16.

5. WorldRoamBench: AutoNavi-led "roaming exam" for interactive world models, 12 models compared

• Sources: Jiqizhixin (Synced) / Sohu Tech
• Date: 2026-09-15

AutoNavi (AMap), with Nanjing University, Tsinghua University, and Peking University, released WorldRoamBench, an evaluation benchmark for interactive world models aimed at the "last mile" of embodied AI deployment. Built on a four-dimensional evaluation system, it uses 1,000+ open-world long-horizon samples to uniformly test 12 world models including Genie 3, Happy Oyster, and LingBot-World, covering first/third-person views across natural, urban, and indoor scenes, with 10–60 second continuous WASD/IJKL interaction simulating real roaming, plus a public leaderboard. Compared to the July version with 600+ test cases, this is an expanded new round. The motivation targets the current state: world models have a key limitation in interactive response consistency — looking realistic is not the same as being controllable.

6. CAS Institute of Industrial AI's Maxwell tops Meta-World leaderboard: instructions parsed into code that commands robots

• Sources: Jiaohui Dian (Xinhua Daily) / Sohu
• Link: https://www.xhby.net
• Date: 2026-09-15

In the latest Meta-World evaluation leaderboard released Sept 15, Maxwell, an embodied AI foundation model developed by the Chinese Academy of Sciences Institute of Industrial Artificial Intelligence (Wu Bin and Wang Shaojiang's team), topped the ranking. Meta-World is a classic manipulation benchmark (cited 2,000+ times; 50 robotic manipulation tasks: grasping, moving, door opening, etc.), and Maxwell's average success rate across four difficulty tiers exceeds all models on the Evo-SOTA comparison list. Maxwell is a multimodal embodied model spanning five modalities — text, image, code, state, action: given a user instruction, the model parses it into code, which then directs the robot — another data point for the "code as intermediate action layer" approach.

7. AI national standards accelerate: 100-standard initiative covers compute, large models, embodied AI

• Source: Cover News
• Link: https://cbgc.scol.com.cn
• Date: 2026-09-15

A batch of AI national standards and technical documents were released as part of the "AI 100 National Standards Special Action Plan," covering compute, large models, and embodied AI; technical standardization documents for AI glasses were published concurrently. Embodied AI standards keep moving up the agenda — following legged robot international standards (ISO scenarios) and real-scenario training initiatives, standardization is shifting from "setting the tone" to "setting the specs."

Editor's Notes

Inference engines are going "embodied-first," adding a fifth route on the deployment-cost axis. Over the past two weeks this newsletter has tracked the cost physics of deployment: FreeToken's scheduling pole, M5 Ultra's unified-memory pole, SSP-BO's representation pole, Mobius's architecture pole. Today APXInf fills a previously missing angle — the edge-engine pole: not saving per-token price or token count, but saving milliseconds and watts per action. The numbers matter: π0.5's flow-matching runs at FP8 41ms on Jetson Thor, 26ms/38Hz after single-step pruning — already real-time-control territory. And after FP8 quantization, LIBERO-10 is 92.2% vs a 92.4% reference: the quantization tax is near zero on the embodied side. This is cross-domain isomorphic to the Qwen3.8 quantization ecosystem finding (daily 08-28): the "structural fidelity" problem of quantized deployment is easier to get right on VLAs than on pure language models — small action heads, vision backbones insensitive to precision. An even more telling signal is the README's first line: "born of the agentic coding era," with CUDA kernels "agentically optimized" — an engine written by agents to run models trained by agents. Harness > LLM, closed-loop on edge silicon. The same day, Agility wrote 9-minute charging and a 10:1 run-charge ratio into its product pitch — deployment economics are being rescaled from "per million tokens" to "per watt, per action," a shift in measurement worth recording more than any single product.

Training grounds become infrastructure: the "centralized pole" of the data-collection spectrum takes shape, facing the "retail pole." The CAICT report gives full coordinates: 70+ built, 40+ in progress, more than half of provinces covered, three clusters, extension into tier-5 cities — the previously missing slot on the data-collection spectrum: state-infrastructure-style centralized collection. It forms the two ends of a spectrum with Mifeng's 20,000 body-free devices going into homes (the retail pole, daily 09-02): on one end, capital-heavy centralized "robot schools"; on the other, gig-economy distributed home data collectors. Both appearing in the same month suggests data production is stratifying like manufacturing. But the report's own admission of an "inverted triangle" (active upstream supply, underdeveloped downstream applications) is honest: training grounds are roads built before the traffic arrives — read alongside Unitree's 250-billion market-cap loss in a month and the picture completes: upstream capital and infrastructure accelerating in (tens of billions in H1, matchmaking events, the seventh training camp) while secondary-market valuations retreat (the G1+ launched with shares below half their peak). The divergence between capacity narratives and valuation narratives (daily 09-14) continues, while the evaluation-layer competition represented by WorldRoamBench and Maxwell is the ruler that determines where that divergence settles — model rankings fluctuate, but "which ruler you measure with" is becoming a more durable asset than "who ranks first."

---

*Editor: Sumu. Data and claims cross-checked across sources; Unitree market-cap figures per Xueqiu and investment media retrospectives; Digit 5 specs from official PR and Forbes; WorldRoamBench case counts per media reports.*