Falling and getting up: what recent research shows about humanoid recovery
Researchers are teaching humanoids to sense a fall early, land on their sturdiest parts and stand up in seconds. Here is what the papers measured, and what they did not.
Von Redaktion von humanoid4rent · veröffentlicht 12. November 2026
Humanoid robots still fall, and the research of 2025 and 2026 treats that as a fact to plan for. Recent papers split the problem into three steps: predicting a fall a fraction of a second early, steering the body so it lands on sturdy parts, and standing up again. Lab results are promising, but they come mostly from one small robot on mostly flat ground.
This article summarises what those papers report, what one manufacturer has done differently, and what it means for anyone placing a humanoid near an audience. Most of the work discussed here is published as preprints on arXiv, which means it has not necessarily been peer reviewed yet.
Why falling is a design problem, not a rare accident
A walking humanoid is balanced on a small area under its feet and constantly corrects itself. Our explainer on how humanoid robots keep their balance covers that in more detail. When a push, a slip or a software error takes the robot beyond what those corrections can handle, it will go down.
The researchers behind one 2026 paper on protective materials state the issue plainly: humanoid deployment is held back by how vulnerable the robots are to falls and by the risk their rigid metal and plastic bodies pose to people and surroundings, according to Soft Responsive Materials Enhance Humanoid Safety (arXiv). Another group, working on unified fall safety, notes in their November 2025 paper that no existing method fully prevents loss of balance.
So the question has shifted from "how do we stop falls" to "what happens in the second before, during and after one".
Step one: noticing the fall early
A protective reaction is only useful if it starts before the robot hits the floor. In SafeFall: Learning Protective Control for Humanoid Robots, submitted in November 2025, the authors add a small neural network that watches the robot's own sensor readings and decides whether a fall has become unavoidable. The protective controller stays switched off during normal walking and only takes over when that predictor fires.
The authors report a warning time of about 410 milliseconds before impact and a false-alarm rate of 0.06% in their evaluation. That is less than half a second, which is enough for a robot to reorient its body but far too short for a human operator to react. In practice, the safe response has to be automatic.
A related line of work tries to avoid the fall altogether by stopping safely. A September 2026 preprint from researchers including Pieter Abbeel and C. Karen Liu, Humanoid Safe Stop via Learned Stoppability Value, trains estimators that judge whether the robot can still come to a controlled halt from its current state. If both estimators agree it can, the robot runs a stop routine; if not, it switches to a damped fallback motion instead.
Step two: falling in a way that limits damage
Once a fall is certain, the goal is to choose how to land. SafeFall uses a reward during training that penalises impacts on fragile parts, such as the camera-equipped head and the hands, and favours landing on the rigid torso. The authors describe the robot twisting its upper body during sideways falls so that it lands on the torso with the head turned upward, and bracing with its arms in forward falls.
On the Unitree G1, SafeFall's authors report that the policy cut peak contact forces by 68.3% and peak joint torques by 78.4%, and avoided collisions with vulnerable components in 99.3% of cases. They also say the method works reliably only on flat or near-flat ground, so stairs and very uneven surfaces remain open problems. Readers can compare the G1's published specifications in our Unitree G1 catalogue entry.
The University of Michigan team behind Unified Humanoid Fall-Safety Policy from a Few Demonstrations combines fall avoidance, a controlled descent and getting up into a single controller. It learns from a handful of human fall-and-recover motions taken from video, plus reinforcement learning, and was tested on a G1 in simulation and on real hardware.
Padding that hardens on impact
Software is only half of the answer. The January 2026 protective-materials paper, from researchers at the National University of Singapore, ETH Zürich and ENGINEAI, fitted the 42 kg ENGINEAI PM01 with protectors made from a non-Newtonian material that stays soft in normal contact and stiffens when struck. Combined with a learned falling policy, the authors report that the robot survived repeated falls, including drops from 3 metres and tumbles down stairs, without hardware damage.
The same authors list limits: long-term durability of the material has not been characterised, and cameras and lidar sensors remain exposed. Padding also says nothing about the person or object the robot lands on, beyond the reduced forces the paper measured on the environment.
Step three: getting back up
Standing up from the floor sounds simple, but it involves many body parts touching the ground at once, which is harder to simulate and learn than walking. Two widely cited papers from February 2025 tackled it on the Unitree G1.
- HumanUP (Learning Getting-Up Policies for Real-World Humanoid Robots, University of Illinois Urbana-Champaign and Simon Fraser University) first lets the robot discover any way to get up in simulation, then refines it into slower, smoother motion. In real-world tests on surfaces including grass, snow, slopes and stone tiles, the authors report a 78.3% success rate for standing up from lying on the back, against 41.7% for the manufacturer-provided controller they compared with. Rolling from face down to face up succeeded 98.3% of the time.
- HoST (Learning Humanoid Standing-up Control across Diverse Postures, Shanghai AI Laboratory with several universities) trains standing-up from varied positions, including against walls and on platforms and slopes. Its authors report a 100% success rate across their laboratory test set-ups and successful stand-ups outdoors on grass, wooden platforms and stone paths.
On timing, HumanUP's authors report that their policy gets the robot up in about 6 seconds, compared with nearly 11 seconds for the built-in controller in their tests. They also observed that the built-in routine leaned heavily on the arms and made those motors heat up significantly, while their policy shifted work to the stronger leg motors. These comparisons reflect the firmware and conditions the researchers used, not necessarily the current factory software.
The Michigan unified controller reports 10 out of 10 successful real-world recoveries on a flat mat and 8 out of 10 on a slippery surface. Small trial counts like these are typical for hardware experiments and should be read as indications rather than reliability figures.
A different approach: sitting down on purpose
Not every company is betting on graceful falls. IEEE Spectrum's Evan Ackerman reported in Digit 5 May Be the First Humanoid Robot Worker That's Truly Safe, published on 15 September 2026, that Agility Robotics designed its 129 kg Digit 5 to avoid, stop or lower itself into a stable seated position when a person comes close, dropping any load first, so that it cannot topple onto someone.
Ackerman describes this as inelegant but effective. The underlying idea is useful for event planners too: the safest response near people is often a controlled stop well before balance is lost, rather than a clever fall.
What the research does not yet show
The papers are encouraging, but several gaps matter for real venues:
- Most results come from one robot, the Unitree G1, which is small and light compared with full-size humanoids. Heavier robots carry more energy into a fall.
- Fall-safety policies are mostly evaluated for damage to the robot. Protection of bystanders is measured indirectly, if at all.
- Stairs, cables, crowded floors and people standing within arm's reach are largely outside the tested conditions.
- Research controllers are not automatically part of the software a rented robot runs. Whether a given unit has any of these capabilities depends on the manufacturer and the supplier.
Our overview of documented humanoid robot incidents shows how falls have played out in public so far.
What this means for event organisers
- Plan for falls, not against them. Even the best-performing research controllers do not prevent every loss of balance. Keep a clear zone around a walking humanoid so that a fall lands on empty floor; our planning guide to safety at public events explains how to size and staff it.
- Assume seconds, not instant recovery. Research robots stand up in roughly 6 to 11 seconds when it works, and some attempts fail. Build a pause into choreography and decide in advance who resets the robot, as covered in what happens if the robot fails on the day.
- Floors matter. Recovery success dropped on slippery, sloped and uneven surfaces in the published tests. Check the surface with the supplier using our venue checklist.
- Ask suppliers concrete questions. Ask what the robot does when it loses balance, whether it can get up by itself, how long that takes, and whether the operator can trigger a controlled stop or sit-down.
- Treat lab figures as a ceiling. Success rates come from small trials in controlled conditions; real performance at a busy event is likely to be lower.
Frequently asked questions
Can a humanoid robot get up by itself after falling?
Many research robots can, and some commercial models have built-in get-up routines. In published tests on the Unitree G1, a 2025 research controller succeeded about 78% of the time from lying on its back across varied surfaces, against about 42% for the built-in routine the researchers compared with. Success depends heavily on the floor, the robot's position and the software version, so ask the supplier what their unit can do.
How long does it take a humanoid robot to stand up after a fall?
In the HumanUP study on the Unitree G1, the researchers' policy stood the robot up in about 6 seconds, while the manufacturer-provided controller took nearly 11 seconds in their tests. These are laboratory figures for a small humanoid. At an event, add time for the operator to check the robot for damage before it continues, and plan for attempts that do not succeed.
Do robots know they are about to fall?
Some research systems do. The SafeFall project trained a predictor that uses the robot's own sensor readings to decide when a fall has become unavoidable, giving roughly 410 milliseconds of warning with a very low false-alarm rate in the authors' evaluation. That is enough time for the robot to change how it lands, but too short for a human operator to intervene.
Does fall-safety research make robots safe to stand next to?
Not on its own. Most fall-safety work measures damage to the robot rather than injuries to bystanders, and tests usually take place on open, flat floors without people nearby. Keeping a clear zone around a walking humanoid, using trained operators and having a controlled stop available remain the practical safeguards at events. This is general information, not legal or safety certification advice.
How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-12. Spotted an error? Write to info@humanoid4rent.com.
Quellen
Die Fakten in diesem Artikel wurden am Tag der Erstellung anhand dieser Seiten geprüft.
- Learning Getting-Up Policies for Real-World Humanoid Robots – arXiv
- Learning Humanoid Standing-up Control across Diverse Postures – arXiv
- SafeFall: Learning Protective Control for Humanoid Robots – arXiv
- Unified Humanoid Fall-Safety Policy from a Few Demonstrations – arXiv
- Soft Responsive Materials Enhance Humanoid Safety – arXiv
- Humanoid Safe Stop via Learned Stoppability Value – arXiv
- Digit 5 May Be the First Humanoid Robot Worker That's Truly Safe – IEEE Spectrum
Mehr aus dem Blog
Möchten Sie bei Ihrem nächsten Event einen Roboter?
Beschreiben Sie einmal Datum, Ort und die Aufgabe des Roboters. Anbieter antworten mit Angeboten, die Sie direkt miteinander vergleichen können.
Anfrage sendenSie arbeiten in der Robotik? Schreiben Sie für unseren Blog. →