Introduction
France has companies working at different layers of physical AI, so they should be compared by the product they actually ship. Mistral AI released Robostral Navigate on July 8, 2026 for language-conditioned robot navigation from a single RGB camera. Wandercraft is deploying its Calvin-40 humanoid in industrial programs and says the robot is running at Renault’s Douai plant. Exotec sells warehouse automation built around its Skypod system and reports deployments across more than 200 customer sites. These products solve different problems, and their published metrics are not interchangeable. This guide keeps company claims attributed, separates deployed systems from announced programs, and leaves undisclosed fleet, reliability and intervention data open.
Key findings
- Mistral AI lists Robostral Navigate as an active specialized model released July 8, 2026.
- Wandercraft says Calvin-40 is operating on Renault’s Douai production line and reports customer commitments for additional industrial programs.
- Exotec reports more than 200 customer sites for its Skypod warehouse system, with current 2026 projects publishing site-level robot counts and throughput targets.
- These systems cover navigation models, humanoid material handling and warehouse automation, so their published metrics should stay tied to each task.
- Public sources still omit many comparable fields, including intervention rates, long-duration failure distributions and full delivered-system costs.
French Physical AI Companies by Technology and Evidence - evidence comparison
The table uses source-backed fields and leaves non-comparable or undisclosed information visible.
| System, category or question | Verified evidence | Interpretation or limitation |
|---|---|---|
| Mistral AI / Robostral Navigate | 8B embodied-navigation model released July 8, 2026; single RGB camera plus language instruction; Mistral reports 76.6% success on unseen R2R-CE. | Model result reported by Mistral; it does not establish factory deployment or manipulation capability. |
| Wandercraft / Calvin-40 | Industrial humanoid program with Renault; Wandercraft says Calvin-40 is running at the Douai plant and reports 12 customers as of September 4, 2026. | Company-reported deployment and customer evidence; comparable uptime and intervention data are not public. |
| Exotec / Skypod | Commercial warehouse automation. Exotec reports more than 200 customer sites; a 2026 Renault Germany project specifies 89 robots and planned throughput above 1,600 containers per hour. | Established warehouse system rather than a general-purpose humanoid or foundation model. |
| Cross-company comparison | Navigation software, humanoid material handling and goods-to-person warehouse automation are different technical categories. | Compare each system against its own task, site, trial conditions and operating metrics. |
The rows use different evidence types. Benchmark success, factory operation and warehouse throughput should not be collapsed into one score.
Evidence classification
- Confirmed by official technical documentation specifications, architecture or access stated by the responsible organization.
- Confirmed by a research paper result reported under a defined experiment, without implying deployment.
- Demonstrated on a real system a physical robot or product performed the documented sequence.
- Company claim without independent verification numerical or operational statement supplied by the company.
- Public evidence insufficient version, control mode, duration, trial count or operating conditions are missing.
Definition and scope
France’s Physical AI ecosystem extends beyond humanoid manufacturers. It includes mobile robots, model infrastructure, teleoperation, navigation, industrial software, tactile sensing and research spin-offs. A foreign company with a Paris office is not classified as French. Laboratories are listed separately, and an integrator is not called a foundation-model developer without evidence.
The comparison uses named products, published versions, customer sites and task-specific measurements. Company figures remain attributed to their publishers, and missing uptime, intervention and fleet data are left undisclosed.
How the complete pipeline works
The companies sit at different layers. Robostral Navigate turns camera observations and language into navigation decisions. Calvin-40 combines locomotion and manipulation for industrial material handling. Skypod uses dedicated warehouse robots, storage racks and control software. Their interfaces, safety systems and operating metrics therefore need separate evaluation.
For deployed systems, useful measurements include task cycle time, human intervention, recovery after faults, uptime and delivered fleet size. For a navigation model, benchmark success should stay attached to the camera setup, instruction set and test environment.
Key systems, products and technical evidence
Mistral’s Robostral Navigate adds a model-layer product for embodied navigation. Wandercraft’s Calvin-40 provides a current French humanoid example tied to automotive production. Exotec provides a separate reference point for commercial warehouse robotics with named customer sites.
The comparison keeps navigation models, humanoids and warehouse systems in separate categories. Benchmark success, customer count, fleet size and throughput stay attached to the original task and test conditions.
Evidence from real systems
Real evidence ranges from commercial robots and factory integration to prototypes. Public videos are classified by task and control mode rather than treated as deployment. Real-system evidence is separated from simulation, internal testing, controlled public demonstrations, pilots and commercial deployment.
Comparison method and engineering tradeoffs
A buyer comparing these systems should keep the unit of evidence fixed: benchmark success for navigation, repeated task execution for humanoids, and site throughput for warehouse automation. Converting those measurements into one score would hide the conditions that produced them.
Failure modes and misleading interpretations
Gaps include undisclosed prices, limited delivery counts, proprietary datasets and demonstrations without intervention logs.
Misleading conclusions about French Physical AI companies often begin with one missing qualifier simulated, teleoperated, target, preorder, internal test or selected attempt.
Practical applications and current maturity
Credible applications include warehouses, inspection, assistive robotics, manipulation research and developer tools. These uses are credible only within the documented task, robot and environment.
Practical use of French Physical AI companies depends on who can diagnose failures and restore service. A laboratory may tolerate manual resets and daily calibration; a factory or home cannot. Support, observability and safe fallback behavior therefore belong in the maturity assessment alongside model or hardware capability.
Open problems and recommendations
Comparable public data are still thin. Useful disclosures would include delivered fleet size, unattended operating hours, human interventions, failed-task counts, recovery behavior and total system cost at named customer sites.
For model vendors, reproducible robot benchmarks and access terms help readers judge capability. For hardware vendors, long runs at named sites with failure and intervention logs make deployment status easier to verify.
Limitations and missing information
- Benchmarks from different robots, versions, environments or control modes are not directly comparable.
- Company-reported metrics are not independently audited unless a separate primary record establishes the same result.
- Code, weights, prices, model versions, APIs and commercial availability can change after publication.
- Long-duration reliability, intervention frequency and complete failure distributions are rarely published.
Conclusion
France has current evidence at several layers of physical AI. Mistral publishes an embodied-navigation model, Wandercraft reports industrial Calvin-40 deployments, and Exotec publishes warehouse-automation projects with named sites and fleet details. Their metrics describe different tasks. Public data remain incomplete on uptime, interventions, failure recovery, delivered fleet size and full operating cost.
Frequently asked questions
Which French companies have current Physical AI or robotics products?
This review includes Mistral AI for embodied navigation models, Wandercraft for industrial humanoids and Exotec for warehouse robotics. They are grouped by product type rather than treated as direct competitors.
What is Robostral Navigate?
Mistral AI describes Robostral Navigate as an 8B embodied-navigation model that uses an RGB camera and a language instruction. Mistral released it on July 8, 2026 and reports 76.6% success on the unseen R2R-CE benchmark.
What deployment evidence is public for Wandercraft and Exotec?
Wandercraft says Calvin-40 is running on Renault’s Douai production line and reported 12 customers on September 4, 2026. Exotec reports more than 200 Skypod customer sites and publishes fleet details for some 2026 warehouse projects.
What information is still missing?
Comparable uptime, intervention rates, failed-task counts, recovery behavior and total delivered-system costs are rarely published across these companies. Those gaps prevent a single evidence-based ranking.
Sources and methodology
Sources for French Physical AI companies were rechecked on September 26, 2026. The review prioritizes official records from Mistral AI, Wandercraft and Exotec, with Reuters used to cross-check Mistral’s July 2026 robotics launch.
For French Physical AI companies, evidence is sorted by test setting and control mode simulation is kept apart from physical trials, teleoperation from autonomous execution, and announced access from a system that can actually be obtained or deployed.
The review does not rank unlike systems. It records product category, evidence type, named deployment or benchmark, and the fields that remain undisclosed.
Related TechniaHQRobot guides
Source review notes
Review date recorded July 11, 2026
Supported by the cited sources
- Real evidence ranges from commercial robots and factory integration to prototypes.
Not confirmed or incomplete
- Gaps include undisclosed prices, limited delivery counts, proprietary datasets and demonstrations without intervention logs.
- Company-reported metrics are not independently audited unless a separate primary record establishes the same result.
- Long-duration reliability, intervention frequency and complete failure distributions are rarely published.
Likely to change quickly
- Prices, model versions, APIs, software access and commercial availability.
- Production, customer pilots, deployments and repository maintenance status.
Share this article
Share the current TechniaHQRobot article page.
Continue reading
Open the latest robotics reporting, Physical AI analysis and hardware notes.