Friction is key to making better robot world models
AI News Desk
·
The Robot Report
··
7 min read
Contactile offers robotic hands and grippers equipped with its tactile sensors.
world models">
Contactile offers robotic hands and grippers equipped with its tactile sensors. | Source: Contactile
World models are the next frontier in robot learning. But, because conditioning on touch remains fundamentally incomplete, they cannot reliably generalize across novel surfaces and objects . A new model class, VμA , proposes to fix that by making friction a first-class input.
The most ambitious direction in robot learning today is the world model: a generalist model of physical reality that a robot can use to predict the consequences of its actions, plan across long horizons, and generalize to situations it has never encountered in training.
If a robot’s internal model of the world is accurate enough, it does not need to memorize every task. Instead, it can reason its way through novel ones.
This is a compelling vision, and the field is moving fast. But deploying world models in real robotic systems requires a step that receives less attention than the models themselves: conditioning. A world model must be conditioned on the robot’s current physical state before it can make useful predictions.
And the quality of that conditioning determines whether the model’s predictions reflect reality, or merely an approximation of it.
Current world model conditioning in robotics relies primarily on two inputs. These are visual observations from cameras, and end-effector position from joint encoders. Some of the most capable systems in the field condition on nothing more than this. For free-space motion tasks, it is often sufficient.
For contact-rich manipulation, it is not enough. The moment a robot touches an object, the information that matters most — what is happening at the interface between fingertip and surface — is invisible to a camera and unresolvable from joint position alone.
In many world model implementations, contact is not even encoded by dedicated sensing . Instead, it is inferred from motor currents in the robot’s joints. A proxy signal several steps removed from the actual contact event reflects the gross mechanical load on the drivetrain rather than the distributed physics at the fingertip.
Some have augmented this with tactile sensors, and the results show meaningful improvement. Tactile conditioning as currently implemented remains incomplete. The majority of tactile sensors encode contact as a deformation image, a pressure map, or at best a three-axis force distribution. These representations tell you that contact is occurring and how hard. They do not tell you the quantity that governs whether the contact will succeed or fail.
In many world model implementations, contact is encoded by motor currents, a proxy several steps removed from the actual physics at the fingertip.
Contactile says its tactile sensors allow adaptive grip in real-time, slip correction, and task execution without customization. | Source: Contactile
The missing quantity is μ, the coefficient of static friction. It is the physical parameter that determines whether a grasped object stays grasped and whether the grip force is sufficient given the surface properties, the load, and the contact geometry. It varies by material pairing, surface condition, temperature, and contamination. It cannot be seen by a camera, inferred from joint currents, or reconstructed from a pressure map.
Without μ in the conditioning signal, a world model is still working from an incomplete description of physical reality. It can learn statistical associations between contact patterns and outcomes in training — but it cannot generalize reliably to novel surfaces, unfamiliar objects, or changed conditions, because the causal variable governing those outcomes is absent from its input space.
A model conditioned on μ is in a qualitatively different position. Friction is not a learned proxy or an inferred signal — it is the physical cause of slip, the direct determinant of required grip force, and the quantity that changes when a surface is wet, worn, coated, or contaminated. Conditioning on μ shifts the world model from statistical pattern-matching towards something closer to causal physical understanding.
Conditioning on μ shifts the world model from statistical pattern-matching towards causal physical understanding — and that changes what the model can generalize to.
Contactile’s PapillArray Tactile Compute Module fills this gap. Unlike conventional tactile sensors that output deformation images or pressure maps, it computes — at the sensor edge, in real time, with no upstream processing latency — the full set of physical quantities required for complete contact conditioning:
These outputs are grounded in contact mechanics, not learned inference. They are reproducible across objects, surfaces, and conditions. And critically, they serve the conditioning stack at both stages of the model lifecycle: enriching training data with causal physical signals, and providing live feedback during deployment.
Robot learning has a productive tradition of naming model classes at the point where a new input modality becomes sufficiently important to warrant its own designation. Vision–language–action (VLA) models crystallized a research direction around natural language as a conditioning signal. Vision–force–action (VFA) models did the same for force.
We propose VμA — vision–μ–action — as the designation for world models that incorporate the full Contactile contact physics stack alongside vision and motor action. The μ symbol is not shorthand for “tactile” or “force.” It is the specific physical quantity that has been absent from every prior model class, and the one whose inclusion makes reliable generalization across contact-rich tasks possible for the first time.
A VμA model ingests per-taxel 3D force, global 3D torque, real-time μ, and per-taxel slip state alongside visual observations, and it outputs motor actions conditioned on the complete physical state of every contact. The term is hardware-agnostic in principle, though in practice, the Contactile PapillArray is currently the only sensor that computes all of these quantities at the edge.
VLA gave robots language. VFA gave robots force. VμA gives robots the physics of contact, and with it, world models that do not just see the physical world, but also genuinely understand what it feels like to interact with it.
VLA and VFA define what robots learn from. World model conditioning defines what they reason with. VμA matters for both, and μ is missing from both.
Training data quality is one half of the VμA value proposition. The other is real-time execution.
The gap between simulation and reality, the sim-to-real problem, is particularly acute for contact-rich tasks, and the reason is precisely the absence of friction physics in simulation. Robots trained in simulation encounter real surfaces with unpredictable μ values and fail in ways their training never prepared them for.
The standard mitigation is domain randomization: Vary the simulated friction parameters widely during training and hope the real world falls somewhere in the distribution.
VμA execution eliminates the need for this approximation. Because the PapillArray measures μ in real time during execution, a VμA-enabled robot does not need to guess the friction properties of what it is holding. It knows. The policy can condition on current μ as a live input, adapting grip force and manipulation strategy to the actual surface, not a simulated distribution of surfaces.
The practical consequences are significant. A VμA robot handling a dry component in the morning and a wet one after a cleaning cycle does not need to be retrained or recalibrated. It measures the changed μ and adjusts.
A robot assembling parts with variable surface finish across a production batch does not accumulate failures as μ drifts. It tracks the drift and compensates. The feedback loop is closed at the contact interface, at the speed of contact, using real physics.
A VμA robot does not need to guess the friction properties of what it is holding. It measures μ in real time and adapts — to the actual surface, not a simulated distribution.
The manipulation ceiling the field has been hitting is not a model size problem. It is a conditioning problem — and specifically, an absence of the right physical data at the right point in the stack. VμA is the framework for solving it.
For the research community, VμA establishes a clear benchmark threshold: a model that does not condition on real-time μ is not a complete contact model, regardless of its architecture or scale. For industry, it defines the sensor specification required to close the manipulation reliability gap that has kept robots out of the most demanding contact-rich applications.
We invite the robotics and machine learning community to adopt VμA as the standard designation for this class of model, and to recognize that the next step forward in manipulation is not a larger model — it is a more complete one.
Heba Khamis is co-founder and CEO of Contactile , a Sydney-based technology company focused on endowing robots with tactile intelligence. She has a Ph.D. in engineering from the University of Sydney.
The post Friction is key to making better robot world models appeared first on The Robot Report .
