Technology
Fraunhofer Tests Unitree G1: 500N Collisions, 1h 49m Runtime, and a New Humanoid Benchmark
Fraunhofer IPA applied a new six-part industrial humanoid benchmark to the Unitree G1 EDU-4, finding good balance and cleanroom potential alongside collision forces above 500 newtons, a Bluetooth vulnerability that has since been fixed, and 1 hour 49 minutes of mixed standing and walking runtime.
Fraunhofer IPA has put the Unitree G1 EDU-4 through a new six-part industrial benchmark, reporting collision forces above 500 newtons, a mixed standing and walking runtime of 1 hour 49 minutes, and potential suitability for ISO Class 5 cleanrooms.
The May 27 release, still moving through robotics trade coverage this week, is notable because it shifts the humanoid conversation from demo clips to third-party measurement. For buyers looking at affordable humanoids, the message is blunt: the hardware is getting real, but safety, security, runtime, and application fit need objective tests before factory rollout.
Key Stats
6
Benchmark Areas
500+ N
Collision Force Finding
1h 49m
Mixed Runtime
2028
Expected ISO Humanoid Safety Standard
How Fraunhofer's Benchmark Works
The Fraunhofer Institute for Manufacturing Engineering and Automation IPA built the benchmark as a modular service for robot makers, software providers, and industrial end users. The intent is not to rate a humanoid on stage presence. It is to test whether a machine can fit a specific production environment with known limits, known risks, and measurable operating behavior.
The benchmark is divided into six categories: technologies and basic capabilities, complex capabilities, cleanroom suitability, functional safety, cybersecurity, and energy efficiency. Each category maps to a practical deployment question. Can the robot move, grip, and carry what the job requires? Can it handle ramps, obstacles, stairs, force precision, and reaction-time demands? Can it operate in a semiconductor, food, pharmaceutical, or other controlled environment without particle emissions causing trouble? Can it collide safely, fail safely, and resist network compromise?
Fraunhofer says the tests are based where possible on recognized industrial standards, including ISO 14644 for cleanroom suitability and ISO 10218 plus ISO/TS 15066 for functional safety. That matters because humanoid-specific safety standards are not expected until ISO 25785-1 arrives around 2028. Until then, factories need a bridge between existing cobot rules and the much harder physics of a legged, whole-body machine.
This is a useful correction to the current market mood. Humanoid companies can show polished clips of walking, dancing, folding clothes, or moving boxes. Those clips rarely disclose collision forces, fallback modes, battery curves, remote-control assumptions, firmware version, network exposure, or particle behavior. Fraunhofer's framework asks the less glamorous questions that decide whether a robot can be insured, approved, serviced, and trusted in a real facility.
Key Insight
A third-party benchmark turns humanoid readiness from a marketing claim into a checklist: safety, runtime, cleanroom fit, cyber hygiene, and task capability all have to survive measurement.
Under the Hood: The Unitree G1 Results
Fraunhofer's first full application used a Unitree G1 EDU-4 with Dex3-1 three-finger hands, delivered in May 2025 and running firmware version 1.04. The G1 is a good first test case because it represents the affordable side of the humanoid market. Unitree's robots are far cheaper and more available than most Western industrial humanoids, which makes them attractive to labs and early adopters, but affordability does not automatically mean production readiness.
AI-generated image
Repeatable measurement exposes the gap between a working platform and a deployable industrial system. Source: AI-generated editorial image.
The positive findings were real. Fraunhofer reported good self-stabilization and said the tested G1 could be suitable for ISO Class 5 cleanrooms, pending application details. That is an important signal for industries where contamination control matters. Semiconductor and pharmaceutical environments do not only ask whether a robot can move. They ask whether the robot sheds particles, whether surfaces can be cleaned, and whether outgassing behavior fits the process.
The negative findings are more commercially urgent. In collision tests, Fraunhofer measured forces exceeding 500 newtons, far above pain thresholds permitted by the relevant collaborative robot standards cited by the institute. The group also identified a critical Bluetooth vulnerability in the tested software version that could allow complete remote control. Fraunhofer says the issue has since been resolved, but the episode shows why humanoid procurement now has to include security review alongside mechanical safety review.
Energy data gave another practical constraint. Fraunhofer reported a maximum operating time of 2 hours 49 minutes while standing still and 1 hour 49 minutes in a mixed standing and walking scenario on one charge. That is not a failure for a research platform, but it changes deployment math. A buyer planning continuous work has to model charger placement, battery swaps, task duration, utilization rate, and idle time before comparing the robot to a fixed automation cell or a wheeled mobile manipulator.
| Test Area | Fraunhofer Finding | Industrial Meaning | Status Signal |
|---|---|---|---|
| Self-stabilization | Good balance behavior reported | Useful for ramps, uneven paths, and recovery | Promising |
| Cleanroom fit | Potential ISO Class 5 suitability | Relevant to semiconductor, pharma, and food uses | Promising |
| Collision forces | Above 500 N in some impacts | Requires guarding, limits, or different work design | Risk |
| Cybersecurity | Bluetooth flaw found, later fixed | Security review is part of robot acceptance testing | Fix required |
| Runtime | 1h 49m mixed standing and walking | Charging strategy affects useful duty cycle | Constrained |
Who's Building the Test Layer
Fraunhofer IPA is one of Europe's most important applied robotics research institutions, with a long record in manufacturing automation, cleanroom technology, and human-robot collaboration. That position gives the benchmark more weight than a vendor comparison chart. The institute is not trying to sell the G1, a rival humanoid, or a robot fleet. It is selling measurement infrastructure to a market that badly needs it.
The institute names Werner Kraus, head of its automation and robotics research division, and Simon Schmidt, senior manager of automated systems, in the release. Schmidt's framing is direct: end users and manufacturers need to look behind the facade built by marketing. That is exactly the pain point for humanoids in 2026. Procurement teams are being asked to evaluate machines that combine robot arms, mobile bases, batteries, AI policies, teleoperation tools, fleet software, and safety systems into one moving body.
Unitree also sits at an important market point. The company has pushed down prices and widened access to legged platforms, which makes its robots popular in labs, demos, and early experiments. The Fraunhofer results do not make the G1 a bad robot. They make it a measured robot. That is healthier for the sector than another unscored video, because both strengths and gaps become visible.
Independent Testing
Fraunhofer guides robots through standardized challenges and reports results against application-relevant criteria.
Comparable Procurement
The goal is to compare humanoids with each other and with proven automation components before capital decisions are made.
What This Means for Industrial Buyers
The benchmark lands at the right time. Humanoid vendors are pushing into logistics, automotive plants, clean manufacturing, retail operations, and research labs. Some deployments are paid and productive. Others are pilots, demonstrations, or data-collection exercises. The buyer problem is that all of those can sound similar in a pitch deck.
Fraunhofer's approach creates a practical filter. If a robot is being considered for a cleanroom, particle emissions and cleanability move to the top of the list. If it is being considered for work near people, collision forces, obstacle detection, and failure behavior matter more than dancing ability. If it connects to a plant network, cybersecurity is no longer optional. If the task requires multi-hour operation, battery life and charging cycles decide whether the robot can finish a shift or only support a lab demo.
The biggest commercial lesson is that humanoid readiness is task-specific. A robot with weak collision limits may still be useful in a fenced area, after speed limits, force limits, and workflow design are changed. A robot with modest runtime may still be useful for intermittent inspection or training data collection. A robot with cleanroom promise may be valuable even if its hands are not ready for fine assembly. The point is to match evidence to the job.
AI-generated image
For factories, the central question is not whether a humanoid can move. It is whether it can fit the safety, runtime, network, and contamination constraints of a specific line. Source: AI-generated editorial image.
Deployment Questions to Ask
• Safety: What peak and sustained collision forces are measured at the actual operating speed?
• Runtime: How many minutes of useful task work happen per charge, not just while standing?
• Security: Which wireless interfaces are enabled, patched, logged, and controlled by policy?
• Environment: Does the robot meet particle, outgassing, cleaning, and temperature requirements?
• Fallback: What happens when balance, perception, grasping, or network connectivity fails?
What's Coming Next
Fraunhofer says it plans to test additional humanoids and build a comparative database. That is the part to watch. A single measured G1 is useful. A multi-robot database using the same test categories would be much more valuable, especially if it lets buyers compare affordable research platforms, premium industrial humanoids, wheeled humanoids, and traditional automation options.
The next 12 months should show whether robot makers embrace outside testing or try to keep performance claims inside controlled demos. The companies most confident in safety, autonomy, and reliability should want neutral benchmarks. The weaker products will benefit from ambiguity. That makes third-party testing a competitive signal before the scores are even public.
Frequently Asked Questions
What did Fraunhofer IPA test?
Fraunhofer tested humanoid robots across six industrial criteria: basic technologies and capabilities, complex capabilities, cleanroom suitability, functional safety, cybersecurity, and energy efficiency. The first full application used a Unitree G1 EDU-4 with Dex3-1 three-finger hands and firmware version 1.04.
Was the Unitree G1 safe for human collaboration?
Not without careful constraints. Fraunhofer reported collision forces above 500 newtons in some tests, which it described as far above pain thresholds permitted by relevant standards. A safe deployment would need application-specific risk assessment, guarding, speed and force limits, or revised workflows.
How long did the tested G1 run on one charge?
Fraunhofer reported 2 hours and 49 minutes while standing still and 1 hour and 49 minutes in a typical mixed scenario involving standing and walking. Those figures are useful for deployment planning because duty cycle, charger placement, and task timing can change the economics of a robot cell.
Why does cleanroom suitability matter for humanoids?
Cleanroom suitability matters in semiconductor, pharmaceutical, food, and other controlled environments where particle emissions or outgassing can damage products or violate process rules. Fraunhofer said the tested G1 showed potential suitability for ISO Class 5 cleanrooms, but suitability still depends on the exact process and application.
Does this mean Unitree's G1 is not useful?
No. The benchmark shows both promise and limits. The G1 demonstrated good self-stabilization and cleanroom potential, while also showing safety, runtime, and cybersecurity issues that need to be managed. That is exactly why third-party testing is useful.
The 12-Month Outlook
Humanoid robotics is entering a measurement phase. The winners will not only be the robots that walk well on video. They will be the platforms that can prove safe contact behavior, stable uptime, clean network posture, maintainable fleets, and task-specific value under conditions that someone else controls.
Fraunhofer's benchmark is one part of that shift, alongside NIST's U.S. baseline work and expected humanoid-specific ISO standards later in the decade. For industrial buyers, this is good news. It gives procurement teams a path to ask for evidence instead of accepting polished clips as proof.
The Bottom Line: Fraunhofer's Unitree G1 test shows that humanoid robots are becoming measurable industrial equipment, and measurement is where the market's hype starts turning into procurement-grade evidence.
Watch the next robots added to Fraunhofer's comparison database. The first broad, neutral scorecards may do more to shape humanoid buying decisions than another month of demo videos.