Skip to main content

Ant Group's Hundreds-of-Millions Bet on Daimon Robotics: Tactile Sensing Becomes the New Robot Entry Point

On August 11, 2026, Daimon Robotics announced a strategic round of several hundred million yuan, led by Ant Group with existing shareholders over-subscribing. This is the first time Ant has extended its embodied-AI investment reach into the tactile-sensing layer. Before this, Ant had already densely placed bets on Unitree, Galbot, Linghearts, Sudu, Wuji, and other body/brain companies — but it had no tactile piece. Daimon completes the puzzle. And Daimon's parallel launch of the world's first Physical Interaction Brain, Daimon-TWM (Tactile-grounded World Model), reveals a deeper truth — tactile sensing is moving from "post-hoc feedback" to "pre-hoc simulation", which may be the real next stop for the robotics track.

Case Background: Capital Is Migrating from Robot Bodies to Tactile Sensing​

Daimon's funding pace is fierce: a hundred-million-yuan A round just two months earlier, with backers already including China Merchants Venture, Lenovo Capital, Inovance Industrial Investment, China Mobile, and China Telecom. After Ant's entry, the capital map further locks down the tactile-sensing layer.

This is not an isolated case. The entire embodied-AI track is seeing a capital migration: away from robot bodies (humanoid, quadruped) and toward the underlying sensing technologies like tactile and torque. The logic behind it:

  • Bodies are oversupplied: Unitree, Agibot, Galbot and other body makers are shipping densely; hardware differentiation is narrowing
  • Tactile is scarce: Globally, very few players can deliver a full stack like Daimon's "tactile sensing device + physical interaction data + evaluation system + model capability"
  • The hardest mile of AGI: Robotic dexterous manipulation is a recognized AI hard problem; tactile data is the breakthrough

Ant's embodied-AI footprint has now expanded from "body—brain" to "body—brain—tactile," and Daimon happens to close the final strategic gap.

Who Is Daimon: CMU → HKUST → Shenzhen, a Decade of Technical Accumulation​

Daimon's roots trace back to Carnegie Mellon University (CMU) — the birthplace of global robotics research. In 1979, Turing Award winner Raj Reddy founded the world's first Robotics Institute there; in 1982, Matthew T. Mason joined and became its director. He is honored as the "Father of Robotic Dexterous Manipulation".

Mason has a core belief: "The dexterity of a hand is mostly about the brain, not the hand." — Hand dexterity depends on the "brain," not the hand itself. This belief shaped Daimon's technical DNA: rather than piling sensors and degrees of freedom onto the hand, build a "brain" that can drive the hand.

Mason's Chinese student is Wang Yu — Mason's first graduated PhD. In 2015, Wang co-founded the HKUST Robotics Institute with Li Zexiang and served as its founding dean, after which he led a long-term attack on high-resolution visuo-tactile sensing. At the end of 2023, Daimon officially began large-scale operations in Shenzhen, with Wang as co-founder and Chief Scientist. His three 90s-born PhD students form the core team:

  • Duan Jiangye (Founder & CEO): PhD from CAS, postdoc at HKUST, MIT AI100 Young Scientist, Hurun U35 Pioneer
  • Yuan Weihao (Chief AI Scientist): former multimodal research expert at Alibaba Tongyi Lab, with experience in both large models and robotic physical operation
  • Du Yipai (R&D & Tactile Lead): core disciple of Prof. Wang Yu in tactile sensing, chief architect of Daimon's monochromatic visuo-tactile sensor

From Mason's manipulation theory, to Wang's tactile technology path, to three 90s-born PhDs respectively driving industrialization, models, and product engineering — Daimon took ten years to turn a scientific proposition about robotic dexterity into a commercially scaling system.

Core Product: The World's First "Physical Interaction Brain" — Daimon-TWM​

In August 2026, Daimon officially launched Daimon-TWM (Tactile-grounded World Model) — the world's first "Physical Interaction Brain" for driving robotic dexterous manipulation.

It shares a similar judgment with T-Rex (recently published by Fei-Fei Li and other top scholars) — that tactile is an independent, high-frequency physical perception channel — but goes further: pushing tactile from "post-hoc feedback" to "pre-hoc simulation", unifying physical cognition, predictive decision-making, and instantaneous control.

Three-Layer Architecture​

Daimon-TWM is a three-layer collaborative system, forming a "slow planning — fast correction" closed-loop physical intelligence anchored on physical cognition:

  1. Physical Cognition Module: Reinforcement-learning-based tactile reasoning base. Builds physical common sense, understands contact state, force, deformation, and material properties.
  2. Predictive Decision Module: World-model-based dynamic prediction hub. Forecasts contact-state evolution and potential failure risk, generating action policies.
  3. Instantaneous Control Module: Tactile-feedback-based high-frequency servo system. Continuously fuses decision commands and real-time tactile feedback, achieving millisecond-level micro-corrections at 100Hz.

Key Technical Metrics​

  • 10B-level parameters
  • Real-time inference on NVIDIA RTX 5090
  • Cross-body, cross-arm deployment potential
  • Introduces learnable unified tactile tokens, building a unified tactile latent space
  • Comprehensive superiority in physical cognition evaluation: in contact-intensive manipulation, average success rate doubles that of π0.5 under no-perturbation conditions, and increases tenfold under perturbation conditions

Typical Scenario: Cleaning Up Broken Glass​

Daimon demonstrated a signature scenario: cleaning up broken glass. Vision struggles to reliably identify transparent, reflective, irregular edges; glass is slippery and fragile; fragments cling to the tablecloth, and lifting them directly might pull the cloth up. Daimon's Physical Interaction Brain confirms the grip via contact, simulates different grasping angles, proactively adjusts to a vertical posture, and uses real-time tactile feedback to continuously correct force and trajectory — finally safely placing fragments into the bin. This is a paradigmatic example of "pre-hoc simulation + instantaneous control" coordination.

Data Infrastructure: The Self-Evolving Flywheel​

Daimon's flywheel logic is "perception generates data, data trains intelligence, evaluation validates capability, deployment drives iteration" — every link has the backing of industrial partners and collaborators:

Perception​

  • Partnership with Lenovo: industrial-grade mass production
  • Pioneer in 10,000-unit shipment of visuo-tactile sensors
  • 80% of dexterous-hand vendors covered

Data​

  • Built the world's largest tactile-omni-modal physical-world dataset: Daimon-Infinity
  • Accumulated hundreds of thousands of hours of tactile-omni-modal physical-world operation data
  • Plans to expand to millions of hours within the year
  • Co-built a "Data Collection to Home" external collection network with China Mobile
  • Deployed the world's first "Embodied Data Collection 5S Store" in Chenzhou, Hunan
  • Phase-1 plan: 1,000 devices, producing 1 million hours of real operation data annually at full capacity
  • First 10,000 hours of Daimon-Infinity open-sourced on Alibaba's ModelScope, with nearly 5 million downloads

Evaluation​

  • Built the industry's first tactile-omni-modal evaluation benchmark for physical interaction capability: RobOmni
  • NVIDIA partnership for ecosystem support

Deployment​

  • Inovance connects industrial automation and precision manufacturing
  • China Merchants connects port and logistics scenarios
  • Already entered supply chains of Inovance and Tesla suppliers

Commercial Track Record: 8 Global Number Ones​

Daimon has cumulatively served 200+ global customers, with over 50 overseas, and delivered to OpenAI, Figure, Physical Intelligence, Skild AI, Meta, BMW, Google DeepMind and other leading global enterprises and institutions. It has achieved global number one in 8 key metrics:

  1. Pioneer in 10,000-unit shipment of visuo-tactile sensors
  2. Pioneer in 80% dexterous-hand vendor coverage
  3. #1 globally in visuo-tactile sensor shipments
  4. #1 globally in visuo-tactile track revenue scale
  5. #1 globally in tactile-omni-modal embodied dataset downloads
  6. Built the world's first large-scale external data collection system
  7. Launched the world's first tactile-omni-modal evaluation benchmark for physical interaction: RobOmni
  8. Launched the world's first cross-body, cross-morphology-deployable "Physical Interaction Brain": Daimon-TWM

The commercial curve is clearly accelerating — "perception generates scale effects, data generates network effects, intelligence lifts the value ceiling" — tactile sensing is becoming a platform-level compound-interest race where entry points, data, intelligence, and standards reinforce each other.

Theoretical Support: Why Tactile Is the New Robot Entry Point​

Behind this case are three theoretical judgments worth remembering.

Judgment 1: Moravec's Paradox. Machines excel at complex logical reasoning like math and code, but struggle to replicate human near-instinctive perception and manipulation. Tactile is exactly the kind of data — within "perception and manipulation" — that is hardest to digitize and scale-collect.

Judgment 2: Fei-Fei Li's T-Rex Empirical Evidence. Tactile is an independent, high-frequency physical perception channel; naively splicing tactile signals into existing VLA architectures doesn't necessarily yield gains, and wrong fusion can even damage original capabilities. Tactile needs dedicated pathways and architectures.

Judgment 3: Mason's 30-Year Unchanged Belief — "The dexterity of a hand is mostly about the brain, not the hand." This means the ceiling of embodied intelligence isn't in mechanical structure, but in the "brain." Whoever masters tactile data and tactile-driven world models, masters the next stop of embodied intelligence.

Case Insights: Why Tactile Becomes the "New Entry Point" for Robotics​

The biggest takeaway from the Daimon case is: embodied-AI competition is shifting from "hardware" to "data + model". Three observations worth noting:

First, the direction of capital migration reflects the direction of industrial evolution. Ant investing all the way from bodies to tactile shows the embodied-AI value chain is migrating to "invisible layers" — once hardware converges, differentiation lies in data, models, and evaluation.

Second, acquiring tactile data is a "capital-intensive business". Daimon took ten years to scale visuo-tactile sensors to 10,000-unit shipments, and build an external collection network of 1,000 devices — behind this is a composite moat of product iteration, engineering yield, and supply-chain coordination. This moat won't be overturned in 1-2 years.

Third, the Physical Interaction Brain is a trinity of "model + hardware + data". Daimon is not pure software, not pure hardware, but weaves all three into a self-evolving flywheel. This kind of "full-stack capability" is hard for single-type players (pure algorithm / pure hardware / pure data companies) to replicate.

Limitations and Open Questions​

Despite impressive results, three open questions remain:

  1. Tactile cost-down curve: Current visuo-tactile sensor unit cost is still high; whether it can be brought down within 2-3 years to a level body makers are willing to standardize on is the key to scale
  2. Cross-vendor data interoperability: Tactile data formats vary by vendor; no unified standard — RobOmni is a Daimon-led standard; whether it becomes industry consensus is still unclear
  3. Distance to physical AGI: Daimon-TWM addresses "contact-intensive tasks," but more complex "long-horizon task planning" and "human-robot collaborative fine manipulation" still need further breakthroughs

Summary​

Ant Group's hundred-million-yuan bet on Daimon is essentially a top-tier industrial-capital vote that confirms tactile as the new entry point for robotics. Daimon's core moat isn't any single technology — it's weaving tactile sensing device + physical interaction data + evaluation benchmark + world model + commercial deployment into a self-evolving flywheel. This full-stack capability is the most scarce in the current embodied-AI track.

Three key takeaways from the case:

  1. Capital direction = industry direction: Giants moving from bodies to tactile signals differentiation is migrating to "invisible layers"
  2. Tactile is a "capital-intensive business": 10,000-unit shipments + 1,000 external devices + ten-year accumulation — hard for new entrants to replicate quickly
  3. Flywheel-style full stack wins: hardware + data + evaluation + model as one — single-point breakthroughs struggle to build lasting moats

References: QbitAI 2026-08-11 report "Ant Group's First Investment in Robot 'Fingertips'", Daimon Robotics official site https://www.dmrobot.com/, Alibaba ModelScope Daimon-Infinity dataset.