Flexion Robotics

Flexion Robotics AG, usually shortened to Flexion, is a Swiss robotics software company that builds a hardware-agnostic autonomy stack, often described as "the brain," for humanoid robots.^[1][2] The company trains its control and motion systems primarily with reinforcement learning in large-scale physics simulation rather than with fleets of human teleoperators, and licenses the resulting software to robot manufacturers instead of building robots of its own.^[2][3] Founded in Zurich in December 2024 by researchers from ETH Zurich and NVIDIA, Flexion raised a $7.35 million seed round and a $50 million Series A within its first year and, in July 2026, demonstrated with Niantic Spatial and NVIDIA a zero-shot real2sim2real navigation pipeline on a physical humanoid.^[1][4][5] In September 2026 the company was ranked sixth in the Top100 Swiss Startup Award and, days later, was named as the humanoid partner in Odyssey's launch of the Odyssey-3 world model, on which Flexion built humanoid control policies.^[20][21][28]

Background and founding

Flexion Robotics AG was entered in the commercial register of the canton of Zurich on December 9, 2024, with the stated purpose of developing, producing, distributing, and maintaining software for controlling humanoid robots.^[6] The company operated quietly through most of 2025 and came out of stealth on November 20, 2025, when it announced its Series A round.^[3][7]

The founding team came out of ETH Zurich's Robotic Systems Lab and NVIDIA's robotics research groups. CEO Nikita Rudin and CTO David Höller worked on GPU-accelerated robot learning at NVIDIA; Prosus Ventures, an investor, credits the pair with pioneering simulation-based robot training at ETH Zurich starting in 2022 and later building Isaac Lab, NVIDIA's open-source robot learning framework.^[8] Humanoids Daily reported that Höller was a research manager at NVIDIA who helped build Isaac Gym, Isaac Lab's predecessor, and that Rudin came from the ETH lab known for the ANYmal quadruped.^[9] The other co-founders are Julian Nubert, who leads perception, and Fabian Tischhauser, who leads robotics hardware; Marco Hutter, professor at ETH Zurich and head of the Robotic Systems Lab, is listed by the company as a co-founder and advisor.^[10] Beyond the founders, Flexion says its early team drew people from ETH Zurich, NVIDIA, Meta, Google, Tesla, and Amazon, with backgrounds in reinforcement learning, control systems, perception, and mechatronics.^[2][3]

Crunchbase News counted 31 employees at the time of the Series A in November 2025; the Unicorner newsletter put headcount at 45 in May 2026; and by September 2026 the company's About page named 58 team members, including people who had worked or done research stays at Meta Reality Labs, NASA JPL, and the Toyota Research Institute.^[7][10][30]

Technology

Flexion positions itself one layer below the robot makers: it does not build bodies, only the software that runs them. The company summarizes this as "We're not building the body. We're building the brain."^[2] The stack is designed to be hardware agnostic, with abstracted interfaces so it can be ported across humanoid platforms from different manufacturers.^[2][11]

As described at the Series A announcement, the stack has three layers:^[2][12]

The Robot Report noted that Flexion deliberately avoids "end-to-end monoliths," arguing that modularity keeps interfaces clean and testable and improves generalization.^[12] The training approach is simulation first: instead of collecting demonstrations through teleoperation, Flexion generates synthetic data in massively parallel physics simulation and relies on sim-to-real transfer to move policies onto hardware, an approach the founders helped establish through Isaac Gym and Isaac Lab.^[8][12] The company's launch material put it bluntly: "No scripts, no tele-op farms, no brittle task-specific logic."^[2][13] Its November 2025 manifesto, "The Hard Part of Robotics is Robotics," argued that teleoperation and scripted choreography "may look smoother today, but they are brittle" and predicted that reinforcement learning in simulation "will outpace every human-led teaching method."^[32]

The simulation-first stance has never been a blanket exclusion of real robot data. The Reflect v0 write-up published the same week described the data strategy as "asymmetric": simulation and synthetic data wherever possible, with real data incorporated "selectively" when it closes specific gaps.^[23] On the TWIML AI Podcast in January 2026, Rudin discussed how reinforcement learning, imitation learning, and teleoperation data are combined to train policies for quadrupeds and humanoids.^[24] By Reflect v1.0 in June 2026 the company stated that its dexterous manipulation used "a VLA trained on teleoperated data" running under its whole-body controller, added that high reliability in that setting was hard to reach on a free-moving humanoid, and said it was working on solving those tasks with reinforcement learning instead.^[11] The September 2026 collaboration with Odyssey, described below, likewise used tens of hours of humanoid teleoperation data on top of a pretrained world model.^[20]

Flexion's public demonstrations have run on hardware from Unitree. Its debut video, released with the Series A in November 2025, showed a Unitree G1 walking through uneven forest terrain in the Swiss Alps, bending down to pick up trash, and dropping it in a bin; Humanoids Daily reported the robot carried a custom perception setup with a ZED stereo camera and NVIDIA Jetson Orin compute.^[9] The Robot Report likewise pictured a Unitree humanoid running Flexion's software.^[12] A lab-tour video published by technologist Andreas Klinger in March 2026, as reported by Humanoids Daily, showed a fleet of 14 humanoids in the Zurich facility, including several Unitree G1 units, PNDbotics' Adam, and the Oli from LimX Dynamics; the Adam was shown placing packages into a shipping box. The report says Flexion trains 4,000 robots in parallel in Isaac Lab and quotes Rudin on the goal of one software stack across the fleet, with the aim of cutting the effort of bringing a new robot to a new task from years to a week.^[25]

Funding

Flexion raised $57.35 million across two rounds in roughly one year, an unusually fast ramp for a European robotics software startup.^[7]

The Series A brought in DST Global Partners, NVentures, NVIDIA's venture capital arm, redalpine, Prosus Ventures, and Moonfire; Flexion's own announcement did not name a lead investor, although EU-Startups and Crowdfund Insider reported the round as led by DST Global Partners.^{[2][3][4][7][12][18]} EU-Startups reported the round as EUR 43 million.^[4] No valuation was disclosed.^[7] The company said the money would go toward growing the Zurich R&D team, scaling compute and robot fleets, opening a US presence, and commercializing with what it called major OEM partners; it has since listed a San Francisco office at 1004 Treat Avenue alongside its Zurich headquarters.^[2][3] Crunchbase News reported that Flexion plans to charge manufacturers an annual per-robot software license.^[7]

Investors framed the bet around training data. Prosus Ventures argued that the binding constraint on humanoids is scalable, high-quality training data, and that Flexion's simulation-first reinforcement learning lets robots accumulate thousands of hours of virtual experience per day, more than teleoperation pipelines can produce.^[8] Startupticker summarized the target markets as industrial settings, logistics, manufacturing, disaster response, and planetary exploration.^[3]

ICRA 2026 live demonstrations

Flexion exhibited at ICRA 2026 in Vienna and said afterwards that it had run 300 fully autonomous live demonstrations over three days: a humanoid going over stairs, grabbing a cardboard box from the floor, setting it down on a table, and heading back to its starting point, with the command layer breaking the task instruction into actions for the motion and control layers and no human involvement during execution.^[26][27][33] The company reported a success rate above 95 percent and framed the data gathered from robots operating in live environments as the way to push that toward "99.99%."^[27] These are company-reported figures from a trade-show setting.

Reflect v1.0

On June 29, 2026, Flexion released Reflect v1.0, which it calls a robotics intelligence platform for "long-horizon" autonomous humanoid work.^[11][15][17] The release video shows a modified Unitree humanoid receiving a single natural-language instruction,^[19] then autonomously retrieving a delivered snack parcel from a ground-floor delivery area, opening doors, taking stairs and an elevator, unpacking the box, and placing the items in a designated drawer, with no human operator.^[11][16]

Architecturally, Reflect v1.0 puts a custom vision-language model at the top as a mission controller that watches the robot's camera feed and continuously replans. Below it, a vision-language-action model trained on real-world data works together with reinforcement-learning skills and a whole-body controller, on top of a runtime that handles communication, process isolation, low-latency inference, logging, and safety checks.^[11][15] On an internal 16-step mission evaluation, Flexion reported that supervised fine-tuning alone completed 38 percent of missions end to end, and that reinforcement-learning fine-tuning raised this to 90 percent; eWeek noted these are company benchmarks without independent validation.^[15][16]

Flexion was unusually direct about limits: the platform operates within bounded task distributions rather than being a general-purpose worker, some objects remain hard to grasp, the mission controller can make wrong visual assumptions, and recovery behaviors cover some but not all failure modes.^[11]

Real2sim2real demonstration with Niantic Spatial and NVIDIA

On July 20, 2026, Niantic Spatial and Flexion published a joint technical post, also carried on Flexion's site, showing what they described as an end-to-end real2sim2real pipeline: an RGB-only navigation policy trained entirely in simulation that transferred zero-shot to a real humanoid robot navigating a real office.^[1][5] NVIDIA's Isaac Sim and Isaac Lab frameworks provided the simulation and training infrastructure.^[5]

The pipeline starts with a single walkthrough of the deployment site using an off-the-shelf 360-degree RGB camera. Niantic Spatial reconstructs the scene as a digital twin: a 3D Gaussian splat (see Gaussian splatting) supplies photorealistic RGB rendering, while MVSAnywhere, a zero-shot multi-view stereo model, generates an aligned collision mesh that holds up even on low-texture surfaces. Both are packaged as a USDZ file in NVIDIA's NuRec volume format and loaded directly into Isaac Sim and Isaac Lab, where Flexion trains navigation policies with massively parallel reinforcement learning, domain randomization, and image encoders trained offline that run identically in training and on the robot.^[1][5] Niantic Spatial says a five-minute 360-degree capture in its Scaniverse app can become a simulation-ready environment this way.^[1] The division of labor: Niantic Spatial handles "faithfully bringing reality into the simulator," Flexion delivers "policies and deployment software tuned to the specific hardware."^[1]

The teams benchmarked four policy variants, each evaluated over 1,024 rollouts with identical spawn and target poses, in two reconstructed offices:^[5]

Only the RGB policy trained inside the Gaussian-splat reconstruction beat the depth baseline; RGB policies trained on generic or synthetic environments did worse than depth in both offices.^[5] The post highlights three failure modes where RGB beat depth sensing: semantic hazards such as a blue mat that is obvious in color but nearly invisible in a depth stream, thin structures like tripods, railings, and cables that stereo depth blurs or misses, and transparent surfaces such as glass doors and windows.^[5] The authors argue the result matters commercially because deploying a policy to a new site has traditionally taken months of on-site adaptation, while this pipeline compresses site capture to a single walkthrough and training to simulation time.^[5] The blog does not name the humanoid platform used in the demonstration.^[1][5]

Stated limitations include that a reconstruction captures a single moment in time with lighting and reflections baked in, and that scenes are static, with dynamic agents left to future work. The teams listed follow-ups including capture from iPhones and fisheye cameras, open-vocabulary semantic labels, scene transformations such as lighting variants and moved furniture, and extending from local navigation to full-task autonomy with language-conditioned behavior.^[5]

Research collaboration with Odyssey on Odyssey-3

On September 15, 2026, Odyssey, a world-model lab founded by Oliver Cameron and Jeff Hawke, announced Odyssey-3 and, in the same post, "a deep research collaboration with Flexion," which it described as "a leader in general-purpose robot intelligence with expertise in reinforcement learning and whole-body control."^[20] Odyssey-3 is an autoregressive diffusion transformer trained on visual observations of the world; to control a physical system, Odyssey attaches an action decoder, a learned output component trained on paired observations and actions, that translates the model's internal representations into that system's controls.^[20]

According to Odyssey, Flexion built its humanoid control policies on Odyssey-3 as a base model and "carried out substantial research and engineering" to do so. "With only tens of hours of humanoid teleoperation data, the resulting system can perform tasks in real time, applying Odyssey-3's pretrained representations through Flexion's work on robot learning and control," the post says. Odyssey adds that in its evaluations these policies "generalize better to environmental changes than the VLA baselines tested, continuing to execute tasks under lighting changes that cause baseline policies to fail."^[20] The post does not name the baseline VLAs, and it publishes no trial counts or success rates for the humanoid tasks; Humanoids Daily made the same observation and called the evidence preliminary.^[20][22] The demonstration videos are captioned with four language instructions:^[20]

Neither Odyssey's post nor Flexion's own statement names the humanoid platform used.^[20][21] Odyssey quotes Rudin, identified as co-founder and CEO of Flexion: "What excites us about Odyssey-3 is the opportunity to build on physical knowledge acquired far beyond a robot's own demonstrations. Combining that foundation with our research in humanoid learning and control opens up exciting possibilities for how quickly robots can acquire useful skills and adapt to unfamiliar situations."^[20]

Flexion's own account is brief. Its news page carried no post on the collaboration as of September 16, 2026; the company's statement is a post on X the same day that quote-tweeted Odyssey's announcement: "We're excited to be partnering with @odysseyml to bring foundation world models into physical applications. Here's a first look at the world action model we've been training. Read more about our collaboration on their blog."^[21] Flexion's choice of the term world action model is its own; Odyssey's post speaks of humanoid control policies built on a foundation world model.^[20][21]

Neither company describes the collaboration as replacing the simulation-trained parts of Flexion's stack described above; Odyssey credits "Flexion's work on robot learning and control" with applying the model's representations on the robot.^[20] As Humanoids Daily noted, the teleoperation figure Odyssey cites (tens of hours) describes the robot-specific data used to adapt a pretrained model, not the total data behind the system; it compares with the "thousands of hours of demonstration, hundreds of operators" per task that Flexion's Series A announcement criticized.^[2][20][22] Flexion had already disclosed in June 2026 that its manipulation VLA was trained on teleoperated data.^[11] What the Odyssey work adds is a pretrained third-party world model as the source of the representations those policies build on.^[20] Odyssey said it planned to release Odyssey-3 publicly in the coming weeks.^[20]

Position in the humanoid market

Flexion is a software-only bet in a market where most prominent humanoid companies, and several robot foundation model startups, pursue vertical integration or general-purpose models trained heavily on teleoperated demonstrations. Humanoids Daily described Flexion's approach as a horizontal software layer, "the Android of humanoids," that hardware manufacturers can license instead of funding their own autonomy R&D.^[9] The company's pitch leans on the founders' robot learning pedigree: the same simulation-based reinforcement learning line of work, from ETH Zurich's legged robots through Isaac Gym and Isaac Lab, that much of the embodied AI field now builds on.^[8][9] NVIDIA sits on both sides of the relationship, as an investor through NVentures and as the supplier of the Isaac simulation stack Flexion trains in.^[5][7]

On September 9, 2026, the Top100 Swiss Startup Award, an annual expert ranking run by Venturelab with UBS, placed Flexion sixth among Swiss startups no more than five years old and gave it the best-ranked robotics startup award, one of six special awards introduced that year.^[28][29][31]

Whether the model works commercially is still open. Flexion's published results, the 16-step mission evaluation, the ICRA demonstrations, and the office navigation benchmarks, are company-run evaluations, and while Flexion's Series A announcement referred to major OEM partners and a May 2026 newsletter profile described it as piloting with major OEM customers, none have been named in its announcements.^{[2][12][16][30]} Flexion's headline demonstrations have run on Unitree hardware, the March 2026 lab tour is the main public evidence of its software running on other manufacturers' robots, and its published quantitative results (the Reflect v1.0 mission evaluation and the office navigation benchmarks) each come from a single platform.^{[1][9][11][12][25]}

See also

• Humanoid robot
• Niantic Spatial
• Isaac Lab
• Sim-to-real transfer
• Reinforcement learning
• Unitree G1
• Gaussian splatting
• Physical AI
• Odyssey (AI lab)
• Odyssey-3
• World model
• World action model

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Closing the Sim2Real Gap for Humanoids. Niantic Spatial, July 20, 2026. nianticspatial.com/...n-humanoid-real2sim-sim2real
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Flexion Raises $50M to Build the Brain of Humanoid Robots at Scale. Flexion, November 20, 2025. flexion.ai/...he-brain-of-humanoid-robots-at-scale
3. ^1 ^2 ^3 ^4 ^5 ^6Flexion raises $50M to build the brain of humanoid robots. Startupticker.ch, November 2025. startupticker.ch/...d-the-brain-of-humanoid-robots
4. ^1 ^2 ^3Zurich's Flexion raises 43 million euros to build the brains behind humanoids. EU-Startups, November 2025. eu-startups.com/...ild-the-brains-behind-humanoids
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Niantic Spatial, Flexion, and NVIDIA: Closing the Sim2Real Gap for Humanoids. Flexion, July 20, 2026. flexion.ai/...osing-the-sim2real-gap-for-humanoids
6. ^Flexion Robotics AG in Zurich. Moneyhouse (Swiss commercial register data). moneyhouse.ch/...flexion-robotics-ag-11605579991
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Exclusive: Founded By Ex-Nvidia Researchers, Flexion Lands $50M To Build The 'Brain' for Humanoid Robots. Crunchbase News, November 20, 2025. news.crunchbase.com/...lding-startup-flexion-raise
8. ^1 ^2 ^3 ^4Prosus Ventures Invests in Flexion Robotics to Build the Intelligence Stack for Humanoid Robots. Prosus, November 26, 2025. prosus.com/...telligence-stack-for-humanoid-robots
9. ^1 ^2 ^3 ^4 ^5A Brain in the Alps: Flexion Raises $50M to Build the "Android" of Humanoids. Humanoids Daily, November 2025. humanoidsdaily.com/...ild-the-android-of-humanoids
10. ^1 ^2About. Flexion Robotics. flexion.ai/about
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Flexion Reflect v1.0 - The Path Towards Long-Horizon Autonomous Humanoid Work. Flexion, June 29, 2026. flexion.ai/...flexion-reflect-v1.0
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7Flexion to use Series A to build sim-to-real, AI systems powering humanoids. The Robot Report, November 26, 2025. therobotreport.com/...d-ai-systems-power-humanoids
13. ^Video: New brain helps humanoid robot handle uneven surfaces with ease. Interesting Engineering, November 21, 2025. interestingengineering.com/...-robot-conduct-tasks
14. (Reference withdrawn: the BeBeez International article previously cited here returns HTTP 404 as of September 16, 2026.)
15. ^1 ^2 ^3Flexion new AI model gives humanoid robots long-horizon autonomy. Interesting Engineering, June 2026. interestingengineering.com/...-in-complex-missions
16. ^1 ^2 ^3Flexion Reflect v1.0 Shows Why Humanoid Robot Software Matters. eWeek, June 29, 2026. eweek.com/...flexion-reflect-robot-software
17. ^Swiss Startup Flexion Robotics Introduces 'Long-Horizon' Autonomous Humanoid Robotics Platform. The AI Insider, June 29, 2026. theaiinsider.tech/...us-humanoid-robotics-platform
18. ^Robotics Software Startup Flexion Raises $50m Series A To Power Humanoid Autonomy. Crowdfund Insider, November 2025. crowdfundinsider.com/...to-power-humanoid-autonomy
19. ^Reflect v1.0 - The Path Towards Long-Horizon Autonomous Humanoid Work (video). Flexion on YouTube, June 2026. youtube.com/watch
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Introducing Odyssey-3: A General-Purpose Physical Intelligence. Odyssey (Oliver Cameron and Jeff Hawke), September 15, 2026. odyssey.systems/introducing-odyssey-3
21. ^1 ^2 ^3 ^4Flexion (@FlexionAI) on X: "We're excited to be partnering with @odysseyml to bring foundation world models into physical applications." September 15, 2026. x.com/...2099941551782674589
22. ^1 ^2Odyssey Unveils Odyssey-3, With Flexion Building Humanoid Control on Its World Model. Humanoids Daily, September 15, 2026. humanoidsdaily.com/...d-control-on-its-world-model
23. ^Flexion Reflect v0 - Towards Generalizable Robot Autonomy. Flexion, November 20, 2025. flexion.ai/...towards-generalizable-robot-autonomy
24. ^Intelligent Robots in 2026: Are We There Yet? with Nikita Rudin. The TWIML AI Podcast, episode 760, January 8, 2026. twimlai.com/...ent-robots-in-2026-are-we-there-yet
25. ^1 ^2Watch: Inside the Zurich Lab Building the "Android" of Humanoids. Humanoids Daily, March 23, 2026. humanoidsdaily.com/...ing-the-android-of-humanoids
26. ^Flexion (@FlexionAI) on X: "We ran 300 fully autonomous live demonstrations over 3 days at ICRA 2026." June 9, 2026. x.com/...2064250246624915952
27. ^1 ^2Flexion (@FlexionAI) on X: "300 demos, with above 95% success rate." June 9, 2026. x.com/...2064250631422914694
28. ^1 ^2Corintis, DeepJudge and PAVE Space make up the podium of the Top100 Swiss Startup Award. Startupticker.ch, September 10, 2026. startupticker.ch/...the-top100-swiss-startup-award
29. ^Top100 Swiss Startup Award 2026. Venturelab, September 9, 2026. top100startups.swiss/award2026
30. ^1 ^2Flexion: The brain every humanoid robot is missing. Unicorner, May 18, 2026. read.unicorner.news/...flexion
31. ^Flexion (@FlexionAI) on X: "We have been named the Hottest Robotics Startup at the TOP 100 Swiss Startup Awards." September 10, 2026. x.com/...2097992226596848011
32. ^The Hard Part of Robotics is Robotics. Flexion, November 19, 2025. flexion.ai/...the-hard-part-of-robotics-is-robotics
33. ^Flexion (@FlexionAI) on X: "Flexion is present at ICRA, stand 89." May 27, 2026. x.com/...2059548119314948521