Inbolt : the French startup giving AI eyes to industrial robots
🔎 Why the autonomous factory relies on embedded vision
The manufacturing industry faces a cruel paradox: robots are everywhere on production lines, but half of the assembly and preparation tasks remain manual. The reason? An industrial robot is blind. It repeats a millimeter-perfect gesture in a perfectly static environment. As soon as a part moves by two centimeters, it stops.
This is exactly the bottleneck that Inbolt is preparing to fill at Automate 2026 in Chicago (June 22-25). The Parisian startup is launching its AI vision-based robotic programming system, capable of moving from a 3D CAD model to execution on the production line in a single flow. No laser recalibration, no custom jigs, no weeks of commissioning.
The concept is radical: you take the digital twin of the part, send it to the robot, and the robot adapts in real time thanks to a 3D camera mounted directly on its arm. The AI does the rest. Inbolt is not selling yet another robot. It is selling the brain and the eyes that existing ones are missing.
Key takeaways
- Inbolt is launching a vision-first robotic programming system at Automate 2026 that closes the loop between the CAD model and factory execution.
- The technology relies on an on-arm 3D camera coupled with an AI vision model that locates the physical part and adjusts the robot's movement in real time.
- The French startup, founded in 2019, raised €15M in Series A (Exor Ventures) for a total of $22.3M raised, and targets an industrial automation market ready for scale.
Recommended tools
| Inbolt Robot Programming | AI vision robot programming, CAD → execution loop | Quote-based (June 2026, check on inbolt.com) | Manufacturers looking to automate without jigs or realignment |
| Hostinger | Web hosting for showcase sites and AI dashboards | €2.99/month (June 2026, check on hostinger.com) | Robotics startups needing fast landing pages |
From CAD to the factory floor: the loop finally closed
Inbolt's core idea is to eliminate the gap between virtual design and the physical reality of the factory. Today, an engineer designs a part in CAD. Then, an integrator spends days, sometimes weeks, programming the robot to interact with that part in real-world conditions.
Inbolt Robot Programming drastically shortens this process. The engineer works directly from the CAD model to build the robot program. They define waypoints, trajectories, approach zones. Then, during execution, the Inbolt Vision Model takes over.
The system scans the real environment, locates the physical part exactly where it is, and adjusts the robotic arm's movement to execute the planned path. This is not simulation. It is real-time, closed-loop adaptation on the production line.
According to Automation.com, this approach covers the full path from virtual commissioning to adaptive robot motion control. It works just as well for static applications as it does for moving lines — a use case that has historically been a nightmare in industrial robotics.
The technology: on-arm 3D vision + hybrid AI
Inbolt's strength does not lie in a single magic component, but in the coherent assembly of three technological building blocks that, together, create a quasi-human behavior.
The onboard 3D camera
Unlike fixed vision systems installed above conveyors, Inbolt mounts its 3D camera directly on the robot's arm. This on-arm architecture offers a decisive advantage: the point of view moves with the robot. The system sees the part from multiple angles during the movement, which drastically improves localization accuracy.
This is what enables human-like bin picking: the robot looks at a bin of bulk parts, identifies the right one, calculates its exact 3D pose, and grabs it. Without any mechanical preparation of the bin, without any prior orientation of the parts.
The AI vision model
The Inbolt Vision Model is an AI model trained specifically on industrial scenes. It doesn't recognize cats and dogs — it recognizes metal parts, plastic components, mechanical sub-assemblies, even when partially occluded or stacked.
Paris-Saclay University, which documented Inbolt's work in a dedicated article, highlights the startup's hybrid AI approach: a combination of classical computer vision methods (geometry, contours, depth) with neural networks for recognition and decision-making.
Adaptive motion control
The final building block: the motion control engine. Once the part is localized, Inbolt doesn't just correct an XYZ offset. It recalculates the robot's complete trajectory to adapt to the actual pose of the part, taking into account the kinematic constraints of the arm, collision zones, and singularities.
This level of integration is what sets Inbolt apart from a simple vision overlay added to a standard robot. The system thinks in terms of movement, not in terms of images.
Human-like bin picking: what it changes in practice
Bin picking — picking parts from a random bin — has been considered the holy grail of industrial robotics for thirty years. Everyone talks about it, very few actually do it in production.
The problem is not technical in itself. A 3D vision system can locate a part. The problem is 99.9% reliability over 8 hours, 300 days a year, with parts that change in reflectivity, carry oil, and are stacked randomly.
Inbolt tackles this problem with its on-arm approach. Rocking Robots reports that at runtime, the system locates the physical part and adjusts the movement to execute the path planned from the CAD model. The key: the CAD model serves as a geometric prior. The AI doesn't have to guess the shape of the part — it already knows it. It just has to find it in the chaos of the bin.
For the operator, the workflow is extremely simplified. You load the CAD model of the new part. You indicate where the bin is located. The robot starts picking. No teach pendant, no manual definition of regions of interest, no tedious calibration at every part changeover.
Automate 2026: the moment of truth for Inbolt
Automate in Chicago is the largest industrial automation trade show in North America. 40,000 visitors, 800 exhibitors. This is where robotic integrators come to find the technologies they will deploy in their customers' factories over the next 12 to 18 months.
According to Robotics and Automation News, Inbolt is officially launching its extended AI-based robotic control and programming capabilities there. The timing is not coincidental.
The industrial robotics market is at a pivotal moment. Manufacturing labor shortages are worsening in all developed countries. Factories must automate an increasingly varied range of tasks, with increasingly shorter production runs. Traditional teach pendant programming has become an economic bottleneck: you cannot spend three weeks programming a robot for a run of 5,000 parts.
Inbolt arrives with a value proposition that aligns exactly with this trend: reducing commissioning time to make automation profitable even on small and medium-sized batches. This is the shift from mass automation to agile automation.
A French startup with a track record
Inbolt is not a lab project that came out of nowhere. Founded in 2019 by Rudy Cohen, Louis Dumas and Albane Dersy, the startup spent seven years refining its technology before this launch at Automate 2026.
Business Insider details the journey: a 3D vision system that scans industrial environments, creates a digital twin, and guides robotic arms autonomously. The article highlights the total funding raised of $22.3M, including a €15M Series A led by Exor Ventures — the investment vehicle of the Agnelli family (Ferrari, Stellantis, Juventus). A strong signal sent to the automotive industry, Inbolt's primary target market.
Maddyness points out that the startup is already commercializing AI vision solutions with a 3D camera and hybrid AI algorithms. The September 2024 funding round was explicitly aimed at accelerating commercial deployment in Europe and preparing for North American expansion — which is precisely materializing at Automate 2026.
The geographical positioning is strategic. Paris is home to one of the world's highest concentrations of robotics and computer vision researchers (Université Paris-Saclay, INRIA, Mines ParisTech). Inbolt draws from this ecosystem while targeting a global market.
Physical AI and manufacturing: where Inbolt fits in
Physical AI — the act of making artificial intelligence models interact with the physical world — has become the new frontier for research and industry. We see this in the race for humanoid robots, where companies like Figure are aiming for factory deployment. But humanoids are not the only path.
Inbolt represents a more pragmatic approach: rather than replacing the existing industrial robot with an experimental humanoid, it grafts intelligence onto the already installed robot fleet. A KUKA, FANUC, or ABB arm can become autonomous with the Inbolt system. No need to change the hardware.
This philosophy is part of a broader trend in applied AI. Open-source models like DeepSeek V4 Pro show that the democratization of AI is no longer solely driven by conversational LLMs. It is driven by specialized models — vision, control, planning — that integrate into industrial pipelines.
The parallel with what is happening in the world of search agents is striking. OpenSeeker-v2 broke the monopoly of industrial search agents by open-sourcing a powerful alternative. Inbolt does the same thing in robotics: it breaks the monopoly of proprietary vision solutions sold by robot manufacturers with excessive margins.
Real-world applications: beyond bin picking
While bin picking is the most visual application, Inbolt targets a broader spectrum of industrial operations.
Assembly with tolerance
In automotive or aerospace assembly, parts are never exactly in the nominal position. Machining tolerances, assembly deformations, and mechanical clearances create deviations. A hard-programmed robot misses its insertion or, worse, damages the parts.
With Inbolt, the robot sees the actual state of the parts before assembly and adapts its movement. Inserting a seal, screwing a cap, or fitting two components together become robust operations despite geometric variations.
Integrated quality control
Since the 3D camera is already there, why not use it to verify the result? Inbolt can validate the pose of a component after insertion, detect a misalignment, or verify the presence of an element. Quality control is no longer a separate station — it is integrated into the robotic gesture.
Moving lines
This is perhaps the most impressive application documented by Automation.com. On a line where parts move along a conveyor, the robot must not only locate the part but also track it in its movement. This requires real-time fusion between vision and motion control. Inbolt handles this case natively, opening the door to the automation of packaging, wrapping, and logistics lines.
The business model: selling the brain, not the brawn
Inbolt does not build robots. It sells a software-hardware system (3D camera + programming software + AI model) that interfaces with existing industrial robots. This is a deliberate business model choice.
The global industrial robot market is dominated by four players (KUKA, FANUC, ABB, Yaskawa) which together account for over 60% of sales. Trying to compete with them on hardware would be suicidal for a startup. On the other hand, the market for embedded intelligence for these robots is fragmented and nascent. Each manufacturer offers its own vision solution, which is often limited and expensive. Inbolt positions itself as an independent, multi-brand compatible third party that brings a superior layer of intelligence.
Pricing is quote-based, which is standard in industrial automation. The model likely relies on a license per robotic cell, with integration and maintenance fees. For an integrator, the calculation is simple: if Inbolt reduces commissioning time from 3 weeks to 3 days, the ROI is immediate even with a significant license cost.
Competition and positioning
The landscape of AI vision for industrial robotics includes several players, but none offers exactly the same combination.
Robot manufacturers (KUKA, ABB, FANUC) have their own integrated vision solutions. But these are proprietary, locked within the manufacturer's ecosystem, and often limited to pre-defined use cases.
Pure vision vendors (Cognex, Keyence, SICK) excel in quality control and inspection, but do not handle motion control. They provide data to the robot, but it is up to the integrator to bridge the gap — which recreates the complexity that Inbolt eliminates.
3D vision startups (Photoneo, Pickit, Mech-Mind) are the most direct competitors. But while most focus on bin picking as an isolated feature, Inbolt is building a complete programming system that integrates vision as a native component of the CAD-to-execution workflow.
What this implies for the underlying AI models
One technical point deserves to be highlighted. The Inbolt Vision Model is not an LLM. It is a specialized vision model, likely based on advanced transformer or CNN-type architectures, trained on millions of images of industrial parts.
That being said, the trend in industrial AI is convergence. Models like Google's Gemini 3.1 Pro (score of 92 on the reference benchmark) or OpenAI's GPT-5.5 (score of 91) demonstrate spatial reasoning and 3D image understanding capabilities that could, eventually, replace or complement specialized vision models.
For now, Inbolt is making the pragmatic choice of a custom model for the recognition of industrial parts. But the boundary between specialized vision and generalist vision is blurring. In 2-3 years, it is conceivable that a model like Anthropic's Claude Opus 4.7 (score of 90) could, with light fine-tuning, accomplish 3D localization tasks currently reserved for dedicated models.
Inbolt is preparing for this convergence by building a system whose value lies not solely in the vision model, but in the complete CAD-vision-motion integration. Even if the vision model becomes a commodity, the execution pipeline remains a durable competitive advantage.
❌ Common mistakes
Mistake 1: Confusing embedded vision and fixed vision
Many integrators think that a 3D camera on the ceiling is enough. The problem: the point of view is fixed, occlusions are frequent, and calibration must be redone with every change to the cell. The solution: Inbolt's on-arm approach moves the point of view with the robot, eliminating blind spots and reducing sensitivity to the cell layout.
Mistake 2: Underestimating vision commissioning time
Integrating a classic 3D vision system often takes longer than expected. Camera calibration, lighting, defining regions of interest, managing reflectivities — every case is a project. The solution: by starting with the CAD model as the priority, Inbolt removes a large part of this manual configuration. The geometry is already known.
Mistake 3: Trying to replace the robot instead of augmenting it
The strategic mistake is thinking that we must wait for humanoid robots to automate non-rigid tasks. The solution: Inbolt proves that a standard industrial arm, augmented with AI vision, can already handle tasks deemed impossible without jigs. No need to wait for Figure or Tesla Optimus.
❓ Frequently Asked Questions
Does Inbolt work with all industrial robots?
Inbolt interfaces with the major manufacturers (KUKA, FANUC, ABB, Yaskawa, Universal Robots). Exact compatibility depends on the controller version and communication protocol. Check with Inbolt for your specific configuration.
What is the typical implementation time?
Inbolt aims for a drastic reduction in commissioning time. The CAD-to-execution approach makes it possible to go from weeks to just a few days for standard bin picking or adaptive assembly applications.
Does the system work with shiny or reflective parts?
Shiny metal parts are a classic challenge in 3D vision. Inbolt uses its hybrid AI (geometry + deep learning) to compensate for reflectivity artifacts. Results depend on the material, but the system is designed for real industrial environments, including oiled and metal parts.
Does Inbolt replace robotic integrators?
No. Inbolt provides the tool, the integrator deploys the solution. The integrator's role evolves — less time spent on teach pendant programming, more time on process optimization and interfacing with the rest of the line.
What is the difference compared to a Cognex or Keyence vision system?
Cognex and Keyence are excellent for inspection and quality control. Inbolt integrates vision into the robot's motion control pipeline. Vision is not a separate step — it is part of the motion.
✅ Conclusion
Inbolt isn't performing magic: it takes a real industrial problem (robot blindness), tackles it with the right tools (onboard 3D vision + hybrid AI + CAD-execution loop), and deploys it at the right time (when the labor shortage makes agile automation urgent). The launch at Automate 2026 will be the first real real-world test against North American integrators. If the system delivers on its promises under real trade show conditions, the Parisian startup will have all the cards in hand to become the de facto standard for onboard robotic intelligence.