Embodied AI Daily Brief · September 16, 2026

Today's Highlights

• Infinigence (Wu Wen Xin Qiong), with Tsinghua University and Shanghai Jiao Tong University, open-sourced the embodied AI edge inference engine APXInf: 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 the robot
• Agility Robotics launched Digit 5: human-collaboration safety as the headline feature, payload raised to ~23kg, 90-minute runtime with 9-minute fast charging
• Xinhua field survey: 70+ embodied AI training grounds are now built and operating nationwide — "robot schools" are becoming new infrastructure
• Amap (Gaode), Nanjing University, Tsinghua, and Peking University released WorldRoamBench: 1000+ long-horizon samples uniformly benchmarking 12 interactive world models

Detailed Coverage

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

• Sources: Sina Tech / GitHub (infinigence/ApxInf)
• Link: https://github.com/infinigence/ApxInf
• Date: 2026-09-15
• Infinigence, with Tsinghua University and Shanghai Jiao Tong University, officially open-sourced APXInf, an embodied AI edge inference engine that declares "embodied AI is highest priority." It optimizes VLA / WAM model inference for edge devices like Jetson Thor / Orin and is officially described as a Rust implementation with zero external dependencies. The first release is highly optimized for Physical Intelligence's π0.5 VLA, supporting BF16 / FP8 / INT8 precision levels. Performance ledger: 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 retention is the more striking result: on the LIBERO-10 benchmark, FP8 quantization achieves 92.2% success and BF16 92.8%, versus the π0.5 reference of 92.4% — essentially no regression. It ships with an OpenPI-compatible websocket server that lets an unmodified openpi-client connect directly. A companion robotics-side repo (RLinf/APXinf-robo, Franka + LIBERO evaluation) and an agentic workflow for porting new models are also included — the README self-describes the engine as "born of the agentic coding era," with CUDA kernels agentically optimized.

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

• Sources: Sina Tech / ifeng Tech / IT Home
• Date: 2026-09-14
• Unitree released the G1+ on September 14, with six upgrades across motion performance, perception/interaction, and intelligence: the neck gains 2 DOFs for flexible head rotation; the motor system is fully upgraded — shoulder/waist peak torque +110% with 72% less heat at equal torque, arm motor torque +43% with 30% less heat, thigh motors 28% less heat at equal torque; and a dual-eye + wide-angle dual-vision system. Standard price is 95,000 RMB incl. tax (EDU research version unpriced); the 9000mAh quick-swap battery remains (~2h rated runtime) with a new DC54.6V/5.7A external power port for continuous operation, plus OTA / WiFi6 / Bluetooth 5.2. That evening, Xiaomi's Lei Jun visited Unitree; with CEO Wang Xingxing accompanying, he watched humanoid martial arts and calligraphy demos and personally sparred with the robot — it staggered several steps on impact but recovered its balance without falling. Per 21st Century Business Herald, Lei Jun's Shunwei Capital is an early Unitree investor. There is a bearish flip side: Unitree's market cap once hit 444.9 billion RMB in its listing month (intraday 1,100 RMB on Aug 19), but fell below 490 RMB on Sept 11, under 200 billion — roughly 250 billion evaporated in a month.

3. Agility Digit 5 unveiled: "safe collaboration with people" as a first-class product feature; run-to-charge ratio from 2:1 to 10:1

• Sources: Agility Robotics official / PR Newswire / Forbes
• Date: 2026-09-15
• Agility Robotics launched the fifth-generation Digit, positioned as "the first humanoid robot engineered for scaled human-collaboration safety." The safety architecture is the core of this generation: proprietary AI algorithms plus new sensor classes continuously monitor nearby people and take autonomous measures — proactive avoidance, stopping motion, or switching to a seated mode; official demos include holding a load steady when a rug is pulled from underfoot. Hardware: per-carry capacity raised from ~16kg to ~23kg (50 lbs); the new battery delivers 90 minutes of runtime and 9-minute fast charging, lifting the run-to-charge ratio from Digit 4's 2:1 to 10:1 — officially claimed to support 20+ hours of daily operation. Forbes previously reported Digit v5 for December shipment. The contrast is pointed: the same week, Unitree competed on motor torque and a 95,000 RMB price, while Agility competed on "daring to let people stand nearby" and charging economics.

4. Xinhua multi-city survey: 70+ embodied AI training grounds operational; "robot schools" become new infrastructure

• Sources: Xinhua / CAICT "Embodied AI Training Ground Research Report (2026)"
• Date: 2026-09-15 / 09-16
• Xinhua reporters' multi-city survey shows embodied AI training grounds — "robot schools" — accelerating nationwide. Per the CAICT report: as of end of June, 70+ training grounds are built and operating, with 40+ more under construction or planned; coverage spans over half of provincial-level regions, forming three major clusters (Yangtze River Delta, Beijing-Tianjin-Hebei, Pearl River Delta) and extending into tier-3/4/5 cities; the supply chain shows a pronounced inverted triangle — active upstream technology supply but underdeveloped downstream applications — with heavy-asset investment characteristics. Same-day Hangzhou developments: on the 16th, the National AI Application Pilot-Testing Base (Embodied Intelligence) was inaugurated in Hangzhou, and the 2026 Global Embodied AI Developer Conference plus the seventh Embodied Intelligence Training Camp opened at Xidian University's Hangzhou Institute (Beijing Humanoid Robot Innovation Center CEO Xiong Youjun in attendance). Per Radar Finance, hundreds of billions of RMB flowed into embodied intelligence in H1, and an industry matchmaking event was held in Hangzhou on Sept 16.

5. WorldRoamBench: Amap-led "roaming exam" for interactive world models, 12 models head-to-head

• Sources: Jiqizhixin (Synced) / Sohu Tech
• Date: 2026-09-15
• Amap (Gaode), with Nanjing University, Tsinghua, and Peking University, released WorldRoamBench, a benchmark for interactive world models positioned as the "last mile" of embodied AI scale deployment. Built on a four-dimensional evaluation framework, it tests 12 world models — including Genie 3, Happy Oyster, and LingBot-World — on 1000+ open-world long-horizon samples spanning first/third-person views and natural, urban, and indoor scenes, simulating real roaming via 10–60 second continuous WASD/IJKL interaction, with a public leaderboard for ongoing tracking. Compared with the initial 600+ test-case version released in July, this is an expanded new round. The motivation targets a real gap: current world models have key limitations 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: Jiaohuidian (Xinhua Daily) / Sohu
• Date: 2026-09-15
• The latest Meta-World evaluation leaderboard (published Sept 15) puts Maxwell, an embodied intelligence model developed by the Chinese Academy of Sciences Institute of Industrial Artificial Intelligence (Wu Bin, Wang Shaojiang team), at the top. Meta-World is a classic manipulation benchmark (cited 2000+ times; 50 robot manipulation tasks: grasping, carrying, door opening, etc.), and Maxwell's average success rate across four difficulty tiers exceeds models on the Evo-SOTA comparison list. Maxwell is a multimodal embodied model spanning five modalities — text, image, code, state, action: given an instruction, the model parses it into code that then directs the robot, another data point for the "code as an intermediate action layer" approach.

7. AI national standards accelerate: hundred-standard action plan covers compute, large models, embodied intelligence

• Source: Cover News
• Date: 2026-09-15
• A batch of AI national standards and technical documents was released as part of the "Hundred AI National Standards Construction Special Action Plan," covering compute, large models, and embodied intelligence; AI glasses standardization documents were published concurrently. Embodied intelligence's position in the standards hierarchy keeps rising — following legged-robot international standards (ISO context) and real-scene training initiatives, standardization is moving from "setting the tone" to "setting the specs."

Editor's Observations

The inference engine layer is going "embodied-first" — a fifth route on the deployment-collapse axis. For two weeks this brief 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. APXInf adds a previously missing angle — the edge-engine pole: not saving per-token price or token count, but the milliseconds and watts per action. The numbers deserve scrutiny: π0.5's flow-matching on Jetson Thor at FP8 41ms, 26ms/38Hz with single-step pruning — already in real-time control territory. FP8 quantized LIBERO is 92.2% vs the 92.4% reference: the quantization tax on the embodied side is nearly zero. This cross-domain mirrors the Qwen3.8 quantization ecosystem conclusion (brief 08-28): the "structural fidelity" problem of quantized deployment is easier to get right on VLAs than on pure language — small action heads, precision-insensitive vision backbones. A subtler signal sits in the README's first line: "born of the agentic coding era," CUDA kernels "agentically optimized" — an engine written by agents itself, running models trained by agents. Harness > LLM closes the loop on edge silicon. The same day, Agility made 9-minute charging and a 10:1 run-charge ratio a product headline — deployment economics' granularity is shifting from "per million tokens" to "per watt · per action," a scale change more worth recording than any single product.

Training grounds as infrastructure: the "centralized pole" of the data-collection spectrum takes shape, facing the "retail pole." The CAICT report gives a full coordinate system: 70+ built, 40+ in construction, over half of provinces, three clusters, extending to tier-5 cities — the previously missing cell in the data-collection spectrum: state-infrastructured centralized collection. It pairs with MiBee's 20,000 body-less devices entering homes (the retail pole, brief 09-02): one end is heavy-asset centralized "robot schools," the other gig-economy distributed "home data collectors." Both appearing in the same month suggests data production is beginning to stratify like manufacturing. But the report's honest self-diagnosis of the "inverted triangle" (active upstream supply, underdeveloped downstream applications) matters: training grounds are roads built first — the traffic hasn't arrived yet. Read alongside Unitree's 250-billion market cap evaporating in a month, the picture completes: upstream capital and infrastructure are accelerating in (hundreds of billions in H1, matchmaking events, the seventh training camp) while secondary-market valuations are accelerating out (G1+ launched with the stock still below half its peak). The script of capacity narratives and valuation narratives parting ways (brief 09-14) continues — and the benchmark-layer competition that WorldRoamBench and Maxwell represent is the ruler for where this divergence finally converges: model rankings rise and fall, but "which ruler you measure with" is becoming a more durable asset than "who ranks first."

---

*Editor: Su Mu · Data and claims cross-checked across sources; Unitree market-cap figures reflect investor media retrospectives, Digit 5 specs come from official PR and Forbes, WorldRoamBench case counts from media reports.*