DEV Community

Shawn
Shawn

Posted on Originally published at archive.futurexon.com

FutureX · Physical AI Daily — Issue 129 (09/24)

Today's highlights

· Qualcomm acquires PickNik, the primary maintainer of MoveIt, and commits to keeping MoveIt open source

· Mercedes-Benz signs a production agreement with Wayve, putting Wayve's AI Driver into production Mercedes vehicles within two years

· Boyin Hechuang orders 2,000 units of humanoid force-controlled arms from FAOU Robotics

· In August, embodied robot tenders in China totaled roughly RMB 550 million, with less than 10% going to real-world deployments like power-grid inspection and industrial inspection

· A surgical policy trained only on the motions of a doctor's hand-held instruments performs appendectomies on live rabbits, succeeding in 3 of 4

Paper progress

Trained only on motion data from a doctor's hand-held instruments, a robot completes bimanual laparoscopic appendectomy on live animals · manipulation

Surgical robot learning usually requires the robot itself to collect demonstrations; this paper bypasses that step entirely. The authors attached a recording device to ordinary laparoscopic instrument shafts, using inertial, ToF, and Hall sensors to reconstruct instrument pose and jaw opening/closing — no external camera needed. After retraining a diffusion policy on 849 demonstrations from 4 live rabbits, the policy was tested on 4 new rabbits with the electrocautery powered on: with the doctor selecting the surgical stage, the policy completed appendectomy on 3 of the 4. Both demonstration datasets were released with the paper.

Dongho Yee et al. · arXiv 2609.25625 source

JAMB: denoising bimanual actions together with future 3D point trajectories · manipulation

Bimanual coordination is hard because when one arm moves, the scene facing the other arm changes too. JAMB denoises bimanual actions and future 3D point trajectories together in the same Transformer, letting the two correct each other during denoising. Average success rate across RoboTwin 2.0's 16 tasks is 83.4%, 23.9 points above the strongest baseline; on 3 real-robot tasks it beats action-only methods by 50 points.

Chuyang Xiao et al. · arXiv 2609.25322 source

PAKT: feeding real-robot RL with demonstrations from the drag-teaching common in industrial settings · manipulation

Industrial assembly needs micron-level precision, success rates above 99%, and a pace that keeps up with human workers. PAKT has operators drag the robot arm through demonstrations via admittance control, then constrains those trajectories with the same kinematic limits used during policy execution, ensuring the robot can actually reproduce them. Across 4 insertion and industrial assembly benchmarks (including data-center compute-tray handling), it cuts cycle time 23%-48% versus HIL-SERL and reduces human interventions 62%-86%.

Lars Johannsmeier, Yashraj Narang · arXiv 2609.25630 source

RoboFollow: how much of a VLA's high score reflects actually understanding the instruction · benchmark

HF 2↑. The authors trace the problem to "low scene entropy": when a scene only permits one task, a policy can score high while barely attending to language at all. RoboFollow gives each training scene multiple action-distinguishable task branches, then applies four escalating perturbation levels, L0-L3. Testing 9 VLA and WAM policies, strong L0 performance largely fails to hold up at L1-L3; remedies like stronger VLM backbones, joint QA training, and CFG don't close the gap. Code and data are open-sourced.

Chang Guo et al. (Shanghai Jiao Tong University AutoLab) · arXiv 2609.25636 source

ME-U0: Li Auto's unified embodied model combining understanding and generation experts · vla

ME-U0 uses a Mixture-of-Transformers to connect an understanding expert with a generation expert: it first predicts the subtask and locates interactable regions, then uses flow matching to jointly generate future RGB, depth, normals, optical flow, and actions. The model is pretrained on about 4,200 hours of robot and egocentric data, reaching 99.0% average success on LIBERO, 82.5% on LIBERO-Plus, plus real-robot validation. Pandaily reports that Li Auto also released ME-Brain-1.0 and an on-device ME-VLM the same day.

Haoran Wen et al. (Li Auto) · arXiv 2609.25627 source · Reported by: Pandaily source

PatchWAM: treating action as a single image patch, no separate action head · world-model

World-action models typically need an extra action head or a dedicated action expert. PatchWAM writes continuous actions into a single patch via a fixed mapping, letting the visual generation backbone jointly predict "how the robot moves" and "what the scene looks like afterward." With full data plus augmented demonstrations, it reaches 91.8% on LIBERO-Plus and 96.12% on RoboTwin 2.0.

Tianheng Wang et al. · arXiv 2609.25961 source

RoboTwin-Phys: VLA and WAM performance drops when mass and friction change · benchmark

Existing simulation benchmarks randomize appearance and layout but largely keep physical parameters fixed. RoboTwin-Phys continuously varies 13 physical properties within realistic ranges and releases over 5,000 expert demonstrations annotated with the actual physical parameters used. Models that hold up well under visual and layout randomization degrade noticeably once physical conditions change.

Jiaqi Zhang et al. · arXiv 2609.26292 source

A humanoid robot braces against the environment with one hand while exerting force with the other · locomotion

When a humanoid robot exerts force during manipulation, the reaction force disrupts whole-body balance. This paper has a Unitree G1 brace one hand against the surrounding environment while the other hand does the work. No human motion data is needed, and the achievable contact force rises from a maximum of 13.5 N unsupported to 60 N, with no retraining needed when the task area changes.

Zongyuan Zhang et al. · arXiv 2609.25486 source

NavSafe-∞: high open-loop scores don't guarantee closed-loop safety · autonomy

NavSafe-∞ is a photorealistic closed-loop driving benchmark covering 28 event categories across 280 scenarios. The authors evaluate 20 end-to-end driving policies and find open-loop gains rarely translate into closed-loop safety. Open-loop RL fine-tuning also produces reward hacking: policies trade away safety margin for driving progress, which cascades into chained safety failures once run closed-loop.

Yuxin Bao et al. · arXiv 2609.26618 source

Other papers today: StrataVLA (injecting 3D geometry into VLA in layers, 98.53% average on LIBERO, up to 88% fewer geometry-model calls, arXiv 2609.26071 source); Silent Sabotage (backdooring LLM controllers via the robot's own action sequences, 2 citations, arXiv 2609.26184 source); StepTrigger (backdooring VLM planners via legged-robot foot-contact states, arXiv 2609.26131 source); DreamStream (a generative closed-loop simulator for end-to-end driving policies, 1 citation, arXiv 2609.26792 source); IndustrialVLA-Bench (a unified-protocol evaluation of 6 open-source VLA/WAM models, arXiv 2609.25562 source); VLAQuantBench (closed-loop evaluation of post-training quantization for VLA, arXiv 2609.25376 source); SafeLoop (adding risk prediction and rollback on top of VLA, arXiv 2609.26313 source); TriWorldBench (evaluating embodied world models via head- and dual-wrist three-view consistency, arXiv 2609.26314 source); HOTICE (whole-body object transport for humanoids in cluttered environments, arXiv 2609.25363 source); Fisheye-VLA (a single fisheye camera replacing a front-facing plus dual wrist camera setup, arXiv 2609.25750 source); Cartesian Hand (a 7-DoF end effector with fully linear-actuated fingers, arXiv 2609.25696 source)

Open source · tools · benchmarks

· Robotiq Contact Core: adaptive gripper maker Robotiq (Canadian gripper manufacturer) has open-sourced a C++ SDK, an official ROS 2 package, and updated Isaac Sim assets, an area previously driven mostly by community effort. All three are the first components of the Contact Core software layer; the company plans to release "embedded intelligence at the point of contact" in Q4 source

· HappyWorld-Bench: a unified evaluation suite for world models, reported by InfoQ source

· Moxianfei (Chinese simulation firm) × Huawei: jointly launched an embodied-intelligence simulation foundation solution source

Funding and deals

Qualcomm acquires PickNik | Acquisition | Amount undisclosed · adjacent

PickNik is the primary maintainer of the open-source motion-planning framework MoveIt, which handles trajectory planning, collision avoidance, and manipulation control for robot arms. Qualcomm plans to integrate MoveIt more smoothly into its Dragonwing platform, starting with the Arduino VENTUNO Q development board, and has committed to continuing open-source development of MoveIt, ROS, and Space ROS. The deal is still subject to customary closing conditions. A chipmaker acquiring the de facto standard software for arm motion planning fills in the layer its own chips are missing between models and real-time control.Source: The Robot Report source; Dealroom source

XIRRA (Xirang Kaiwu) | Seed + Angel rounds | Several hundred million RMB · world-model

The company closed two rounds within two months of its founding, led by Dunhong Asset Management, with Huakong Fund, Sanhua Holdings (Chinese auto-parts group), and Yinxingu Capital participating. Founder and CEO Li Yin previously served as CTO of large models at Huawei Cloud; co-founder Zhang Hanwang is a professor at Nanyang Technological University. The team is not following the "VLM + action head" approach; instead it is pretraining a native physical foundation model, LPM, from the ground up, with an autoregressive component handling causal reasoning and diffusion handling rendering. The company says it ranked first in trajectory accuracy on the WorldArena 2.0 video-quality track. 36Kr put its valuation at RMB 3.5 billion.⚠️ Single-source claimSource: Zhidongxi source; 36Kr source

ORCAUBOAT | Series C | Nearly RMB 300 million raised this year · autonomy

The Series C closed in July, led by Xigao Investment, with Yuanhe Puhua and Sanyuan Capital participating; early-stage shareholders include Li Zexiang, Gao Bingqiang, and the Alibaba Entrepreneurs Fund. The company builds unmanned surface vessels and shipboard autonomous-driving systems, and says it now operates over a thousand surface robots across more than 10 countries, with more than 2 million kilometers of autonomous navigation and four China Classification Society type-approval certificates.⚠️ Company claimSource: PEdaily source; 36Kr source

Midcentury | Seed round | $15 million · adjacent

Robot data company Midcentury has come out of stealth. It doesn't build robots — it sells training data and simulation tools. The company says it has collected over 2 million hours of first-person human video across more than 50 environment types and over 20,000 tasks, with hands visible in 90% of the footage, annotated with 3D hand and body pose; it also has a simulation platform called Matrix.Source: BlockBeats source

Shiwei Sensing | New round | Tens of millions of RMB · hardware

The company builds robot sensing and transmission components; Zhidongxi describes it as "building a nervous system for robots."Source: Zhidongxi source

Commercialization and deployment

Aurora's 200 driverless trucks are fully allocated to customers by year-end; Hirschbach plans to own 500 of its own · autonomy

At its September 23 analyst and investor day, Aurora said all 200 driverless trucks operating under TaaS agreements by year-end are already allocated to customers, and that the first major customer under its other model, DaaS, Hirschbach, plans to purchase and operate 500 trucks of its own, with deliveries starting in 2027. Cumulative driverless miles to date exceed 500,000 miles, and the company's target is 30,000 trucks by 2030; its stock fell 6.7% that day.⚠️ Plan-stage claimSource: Business Wire via Yahoo Finance source; Traders Union source

Boyin Hechuang orders 2,000 humanoid force-controlled arms from FAOU Robotics · industrial

The two companies held a signing ceremony. Boyin Hechuang was co-incubated by Bosch's Boyuan Capital and Galbot (Chinese general-purpose robot startup); its BW10-Lite and BW10 models both use FAOU's custom 7-DoF joint modules, with joints reaching up to 180°/s. BW10-Lite has already been validated on automotive electronics production lines for tote picking, depalletizing, and machine tending. Boyin Hechuang had previously signed a strategic procurement agreement with Dijie Industrial for nearly 2,000 units of the BW10 series.Source: Gasgoo source

China's August embodied-AI tenders total RMB 550 million, with roughly 60% going to data-collection and training centers · adjacent

Citing GGII data, PANews reports 168 publicly disclosed embodied-AI robot procurement awards in China in August, totaling roughly RMB 550 million, with only 7 orders at the ten-million-RMB tier or above. Unitree products appeared in 55 procurement orders, ranking first by unit count with 136 mid-to-full-size humanoid and quadruped robots. The largest single order was a RMB 122 million project for a data-collection and training center in Taiyuan, Shanxi. By value, data-collection and training centers account for roughly 60%; by project count, research and education applications account for nearly 80%, while real-world deployments — power-grid inspection, industrial inspection, cultural tourism, firefighting — account for less than 10%. This lines up with the skepticism CEO Shao Tianlan of Mech-Mind voiced last week about circular transactions at data-collection centers — pointing to the same underlying question of where the money is actually going.Source: PANews via TradingView source; CryptoRank source

Kargobot's L4 truck fleet approaches 100 units; company claims 10%-18% higher gross margin for mixed human-autonomous fleets · autonomy

On September 22, Kargobot (Chinese autonomous-trucking startup) launched its "scaled AI transport network" at Qipanjing, Ordos, with its L4 truck fleet approaching 100 units and routes now extending from Inner Mongolia into Shaanxi, Shanxi, Gansu, and Xinjiang. Its model has one driver overseeing multiple trucks. COO Li Xiaoxiao said "our unit economics aren't purely modeled — they're measured over a period of real operation," and claimed gross margin is 10%-18% higher than with human drivers, after accounting for depreciation, maintenance, and insurance.⚠️ Company claimSource: National Business Daily via Sina Finance source

UBTECH's U1: 13,361 units ordered, but only 1,500-2,000 expected to ship this year · humanoid

U1 deliveries began September 16 (previously reported). The first customers are auto dealership groups Yongda Group and Baoxin Group, along with South Korea's Galaxy Robot Land; Baoxin's agreement covers both "retail and leasing." At the June launch event, UBTECH (Chinese humanoid robot maker) announced 13,361 units in all-channel orders, with chairman Zhou Jian saying at the time the company would aim to deliver all of them within the year. At the late-August earnings call, CFO Zhang Ju revised this year's deliverable volume to roughly 1,500 to 2,000 units, attributing the shortfall to production bottlenecks in bionic skin and micro servo motors. The presale deposit is RMB 3,000, fully refundable before July 15; the company has not disclosed the proportion of final payments collected.Source: Shizhibang via Sina Finance source

Tesla says a Cybercab drive unit rolls off the line every 10 seconds · autonomy

The figure comes from a September 22 post on Tesla's recruiting account, which also called Cybercab "the most energy-efficient EV ever certified and produced." This refers to drive-unit production cadence, not overall vehicle output.⚠️ Company claimSource: TeslaNorth source

Wien Energie adds 4 more Spot units, covering three power plants · industrial

Starting in 2027, Wien Energie will deploy new Spot units at its Simmering and Donaustadt power plants and its Flötzersteig waste incineration plant, 4 units in total. The first unit, moved to routine inspection duty in 2023, has completed over 4,000 inspections and walked roughly 1,600 kilometers. The team trains its own custom vision models for tasks like reading gauges and checking oil levels.Source: Boston Dynamics source

Industry developments

Mercedes-Benz signs production agreement with Wayve, targeting deployment within two years · autonomy

This marks the first time Wayve's end-to-end AI Driver will go into a premium automaker's production vehicles. It will integrate with Mercedes-Benz's own hardware, MB.OS, and mapping interfaces, providing point-to-point assisted driving for city and highway driving. Mercedes-Benz CTO Jörg Burzer called it "the global first integration of the Wayve AI Driver in the premium segment." The announcement did not specify an SAE level or particular models, giving only "within the next two years" as the timeline. Mercedes-Benz is now running three parallel tracks: its own L3 Drive Pilot, an Nvidia-based L2+ system, and Wayve. Wayve closed a $1.5 billion Series D in February at an $8.6 billion valuation, with Mercedes-Benz, Nissan, and Stellantis among the investors.Source: Electrek source

Toyota plans to spend $6.42 billion annually on robotics starting in 2028 · industrial

Per Ars Technica, Toyota's long-term plan is to deploy roughly 400,000 robots across the group — 150,000 at vehicle assembly plants and 250,000 at parts and materials plants, counting both humanoid and non-humanoid units. Toyota has not disclosed how many are actually deployed today or what share is humanoid. Its wheeled humanoid, Ereii, is already on the assembly line, with workers wearing grippers that mimic its finger structure to teach it tasks. Senior Managing Officer Hiroki Nakajima said the goal is "robots coexisting with people, not replacing them."⚠️ Plan-stage claimSource: Digital Today source

Foxconn and IAT form joint venture Beijing Futekang, a robot contract-manufacturing platform · humanoid

Launched September 21 in Beijing's Yizhuang district, the venture offers one-stop "turnkey" services from R&D through pilot production to mass manufacturing; it signed cooperation agreements the same day with Songyan Dongli (Chinese humanoid robot maker), Shanghai Kepler, RoboParty, and others. Robot companies commonly get stuck between prototype and mass production, and Futekang is positioned to do contract manufacturing for exactly that stage.Source: 36Kr source

Global quadruped robot shipments reached about 35,000 units in H1, with 40% used in industrial applications · industrial

Citing Counterpoint data, Yicai reports Unitree leads total shipments with a 37% share; within the industrial segment, AgiQuad (spun off from Agibot, Chinese humanoid startup) holds nearly 28% and DEEP Robotics 26%. Mid-to-large heavy-duty quadrupeds account for roughly a fifth of shipments. Counterpoint expects full-year shipments to exceed 90,000 units for the first time.Source: Yicai source

China Southern Power Grid signs strategic cooperation framework agreement with Unitree · humanoid

Signed September 22; public information does not specify procurement volumes or use cases.Source: Guandian.cn source

Musk: possibly 100 billion humanoid robots in 20 years · humanoid

In an interview with China Central Television, he also said there would be at least 1 billion within 10 years.⚠️ Remarks made in a speech/interviewSource: Cailianshe source

Hardware · supply chain

· Unitree Dex5-S dexterous hand: 22 degrees of freedom, 1:1 human hand size, weighs 620 grams, 1 kg payload per hand (2 kg max), control frequency up to 1000Hz over Ethernet or USB, palm camera can connect directly, starting price RMB 39,900 source; source

· Tashan Technology: shipping tactile sensors at tens of thousands of units per month; the company says H1 orders already exceed four times all of last year's, and it has completed its shareholding restructuring source

· Hygon 1000-series CPU: an x86 embedded processor aimed at industrial robots and edge AI source

· Nvidia Jetson Orin Nano 2: a new module aimed at entry-level robotics and edge AI source

Top comments (0)