The humanoid AI that finally carries a box across a room may have just taken its first practical step — without ever being shown the complete job. A research team at Seoul National University built RISE, a vision-language-action framework that learns walking and object handling from separately collected demonstrations and stitches the two skills into seamless, multistep tasks.
No one demonstrated the transitions. No one had to. This article unpacks what the humanoid AI result from Seoul actually proves, why walking and grabbing have haunted robot learning for years, how RISE’s roster of specialist “experts” hands work off mid-task, and where this humanoid AI technology will show up first — warehouses, factories, and eventually hospitals and homes.
Table of Contents
What the RISE Humanoid AI Actually Does
One model, two skill families
Professor Songhwai Oh of Seoul National University’s Department of Electrical and Computer Engineering led the team behind RISE — Responsibility-Induced Specialized Experts — a vision-language-action (VLA) framework for humanoid robots. The pitch is deceptively simple: let a humanoid AI learn locomotion and object manipulation from datasets collected independently, then perform tasks that alternate between the two, in sequence, on a real robot.
The team validated the humanoid AI framework in simulation and on a physical Unitree G1 fitted with two Dex3-1 dexterous hands — 43 controllable degrees of freedom in total. The research was accepted at IROS 2026, one of robotics’ two flagship conferences.
Why the demo matters more than the specs
Plenty of robots walk. Plenty of robots grab. The humanoid AI problem lives in the seam: knowing when walking should stop and reaching should start. Conventional systems only learn that timing by watching complete multistep demonstrations, which are slow, expensive, and brittle to record — one fumbled grasp and the whole take is wasted.
RISE deletes that requirement. The transitions are never demonstrated; the robot learns when to switch by listening to its own internal specialist weights change.
The RISE humanoid AI at a glance
| Item | Detail |
|---|---|
| System | RISE (Responsibility-Induced Specialized Experts), a vision-language-action framework |
| Team | Prof. Songhwai Oh, Seoul National University, Dept. of Electrical and Computer Engineering |
| Hardware | Unitree G1 humanoid with two Dex3-1 hands; 43 controllable degrees of freedom |
| Data | Separately collected locomotion and manipulation demonstrations — no combined task demos |
| Venue | IEEE/RSJ IROS 2026 |
| Headline result | 87.3% average success in simulation vs 73.6% for a GR00T-based baseline |
How the demo actually plays out
In the video released with the paper, the humanoid walks to a table, pauses, reaches for an apple, carries it, places it, then walks again for a second object — one continuous run, with no remote scripting of the transitions. The instruction is natural language; the camera feed is the robot’s own. Nothing tells the system “you may now reach.” That quiet moment, where the robot decides for itself that walking is done, is the entire contribution compressed into one gesture — and the reason this humanoid AI result landed at a flagship conference rather than a workshop.
Why Walking and Grabbing Rarely Travel Together
Two bodies of work in one skeleton
Vision-language-action models have become the humanoid AI architecture of choice because they fuse camera input, natural-language instructions, and physical action in one pipeline. In theory, a single VLA should handle “pick up the bottle and carry it to the table.” In practice, locomotion and manipulation involve substantially different joints, dynamics, and failure modes, and one model struggling to master both usually masters neither.
The demonstration trap
The bottleneck is not data — it is the wrong kind of data. Single behaviors are easy to collect: a teleoperator walks the robot across a mat fifty times, or trains it to pick an apple from a bin. Multistep sequences are another story. If the robot drops the apple at step three of five, the entire demonstration is spoiled, and the collector starts over. Teams that tried this directly found collection costs rising with every added stage.
Worse, isolated behavior data teaches nothing about transitions. Walking-then-grabbing data is where a humanoid AI learns what “finished walking” looks like from the inside — and how to begin reaching without being told.
Why the seam is where robots fail
Industry already sees this failure pattern in the field. Whole-body control demos that dazzle in a lab corner collapse when tasks chain together across rooms, as covered in reports on humanoid robots navigating narrow gaps with whole-body AI control. Motion-control breakthroughs — like humanoids learning to sprint and perform spin kicks from human motion data — solve single behaviors brilliantly, while the handoff between behaviors stays unsolved.
Why bigger models don’t fix this
Scale is the default answer in AI, and the default answer fails here for accounting reasons. Ten billion parameters of walking data still contain zero examples of the handoff nobody recorded. Throwing compute at a missing supervision signal buys a more confident gap, not a bridge. The humanoid AI community spent years watching bigger single-behavior models post impressive clips while real tasks stayed choreographed behind the scenes. RISE’s contribution is orthogonal to scale: it changes what data you need, so every future round of scaling multiplies a cheaper foundation.
The stakes beyond one robot
This is why a humanoid AI result like RISE gets read so closely: the transition problem caps every downstream application. Until robots chain skills reliably, they stay demos. The moment chaining works with cheap, separate datasets, the economics of training an entire industry of robots change overnight.
Inside RISE: Specialists and a Meta-Controller
The architecture in one paragraph
RISE combines three parts: a shared visual-language perception model that reads the scene and the instruction, several action experts that propose motor commands, and a meta-controller that decides how much each expert’s proposal counts right now. Perception is shared; action is plural; control is delegated.
Responsibility: earned, not assigned
During training, RISE scores how accurately each expert predicts the current action and calls that score its “responsibility.” Experts with higher responsibility get more influence over the output for that data. Nothing in the setup says “you walk, you grab” — yet over training, experts naturally drift into specializing as locomotion experts or manipulation experts, because that is what the scoreboards reward.
Transitions read from the dials
The elegant trick is what happens at execution time. The meta-controller’s expert weights shift with the behavior in progress, and RISE reads those weight changes as a signal: when the balance swings hard enough, the current task has ended and the next predefined task begins. The humanoid AI effectively notices its own change of mind.
What the robot sees in the moment
At execution time, the shared perception model keeps reading the scene while the meta-controller reweights the experts on every control tick. The team describes watching those weights like gauges: locomotion experts swell while the robot strides, manipulation experts spike as the hand approaches an object. A sudden swing is treated as task completion — the internal equivalent of a person deciding they have finally arrived at the shelf. The humanoid AI never receives an explicit command to switch; it reads its own changing mind and moves on.
Why this beats a hand-written scheduler
A conventional alternative is a finite-state scheduler: hand-code rules for when walking ends and reaching starts. Rules break on the first slightly displaced box, slightly narrow doorway, slightly different bottle. RISE’s transition signal is learned from the same data as the skills themselves, so it inherits their variability.
The team reports that this is exactly why reliability improved as tasks grew longer. Where scripted handoffs accumulate errors stage after stage, the learned handoff held up, and the humanoid AI stayed on plan through five-step tasks on real hardware.
The Results That Make This Humanoid AI Stand Out
Simulation: a double-digit win
In simulation, tasks chained two to four stages — walking to a target, picking apples, bottles, or cups, carrying them, placing them — RISE averaged an 87.3% success rate. The baseline, a single-expert VLA built on top of NVIDIA’s GR00T N1.5 architecture, managed 73.6%. A fourteen-point margin in a benchmark family where single-digit gains are the norm is a decisive humanoid AI result.
Real world: the margin holds
On the physical Unitree G1, tasks stretched to five stages, including pick-and-place subtasks. RISE finished 50.0% of runs; the baseline managed 30.0%. Physical hardware compresses claims and punishes fragility, and RISE kept a twenty-point lead on real floors. Absolute numbers stay modest; the delta is the argument.
What still trips the robot up
Failures cluster where physics punishes hesitation: missed catches in mid-air, objects nudged off course while turning, placements that knock neighboring items. The 50% real-world ceiling mostly reflects manipulation imprecision under whole-body motion, not confused transitions — the handoffs themselves rarely failed. That diagnosis matters for the roadmap: the transition problem this paper set out to solve is close to solved, while dexterity under locomotion disturbance remains the binding constraint for this humanoid AI and for nearly every competitor’s.
The comparison table
| Setting | Task length | RISE | Single-expert baseline |
|---|---|---|---|
| Simulation | 2-4 stages | 87.3% | 73.6% |
| Physical Unitree G1 | up to 5 stages | 50.0% | 30.0% |
| Demo data required | Any | Separate locomotion + manipulation sets | Separate sets (no transition learning) |
| Transition handling | Any | Learned from expert weights | Not addressed |
Read the failure modes, not just the bars
Two other patterns in the results deserve attention. First, RISE’s advantage grew with task length: the longer the chain, the wider the gap over the baseline, exactly what a learned transition signal should produce. Second, on single-stage tasks the two systems drew — consistent with the claim that RISE adds nothing when no chaining is required.
That combination is the signature of a targeted fix rather than a generally better model: same competence on isolated skills, dramatically better behavior once a task requires a humanoid AI to change its mind mid-run.
Why the Data Math Matters So Much
The hidden cost of “just collect more data”
Modern robot AI runs on demonstrations, and this humanoid AI result attacks the line item that makes demonstrations expensive: wasted takes. A single-behavior dataset is cheap and clean. A complete task-sequence dataset multiplies the probability of a spoiled take with every added stage, so its cost grows roughly exponentially while its value grows only linearly.
RISE inverts that curve: collect each behavior in isolation, at industrial cleanliness, and let the architecture recover the transitions. For any lab budgeting a humanoid AI program, that is a procurement decision, not just a research detail.
A market pricing its own data problems
The economics are already moving. A training-data broker reached a near-$500M valuation on the back of the rush for robot training data, and China’s data regulator has begun a standards push for embodied AI data. Frameworks like RISE that shrink demonstration requirements change what that data supply chain even needs to sell.
What standardization can and cannot do
Data standards will make robot datasets interchangeable; they will not make bad datasets sufficient. The RISE result suggests the smarter question is architectural: which frameworks extract the most competence from the cheapest, most reliably collected slices of experience? That question now has a published, peer-reviewed answer.
The quiet simulation-to-real story
Beneath the headline numbers sits a quieter result worth more to engineers than either win. RISE’s drop from simulation to physical hardware — 87.3% down to 50.0% — tracked the baseline’s own decline instead of exaggerating it. Frameworks routinely look brilliant in simulation and collapse on real robots; this one kept its relative advantage on physical hardware. For practitioners deciding whether published humanoid AI techniques transfer to their own fleet, margins that survive the sim-to-real gap are the most portable evidence a paper can offer.
The open-source current
The humanoid AI toolchain is also moving toward shared building blocks — NVIDIA’s ecosystem keeps pushing agentic, open-source robotics development with Isaac ROS 5.0 — which means techniques like responsibility-weighted experts can spread quickly once published. Expect imitations and extensions before the conference circuit finishes printing the proceedings.
Where This Humanoid AI Will Work First
Logistics: the obvious first job
The research team itself points to work where locomotion and manipulation already interleave: moving goods through logistics centers and shuttling parts and tools across manufacturing floors. These are precisely the environments where multistep walk-carry-place tasks repeat thousands of times a shift — and where a 20-point reliability gain converts into money.
The commercial field is crowding in already. Companies building warehouse orchestration are discovering that getting robots and humans on the same page is the real product, and transition-competent humanoids slot straight into that roadmap.
Hospitals, caregiving, homes — later
The paper’s longer list reads like a demographic forecast: transporting supplies in hospitals, assisting with caregiving, organizing objects in homes. Each demands environment adaptability and safety margins the current humanoid AI has not yet demonstrated. Seoul National University’s team is explicit that applications expand only “as adaptability to diverse environments and safety are further improved.”
Why investors watch the seams
Physical-AI investors argue the category is bigger than humanoid robots alone, urging money beyond humanoid hardware into the wider physical AI stack. Results like RISE explain the instinct: the scarce asset is not legs or hands — it is software that connects them.
What workers should expect
On the warehouse floor, transition-competent humanoids change task design before they change headcount. A robot that can walk-carry-place in a single run absorbs the walking dead-time that human workers currently cover for machines — and it creates new supervisory jobs in exception handling, reload stations, and flow selection. Factories piloting arms on fixed pedestals will finally see a concrete argument for giving them legs, because a mobile manipulator only pays off when chaining behaviors stops requiring a fresh demonstration budget per task.
Why insurance and safety teams care
A robot that changes behavior mid-task raises questions insurers have never priced. When does a walking robot become a reaching robot near a person, and who owns the two meters in between? Learned transitions at least make the question answerable: the meta-controller’s weights record when the system believed it was switching, leaving an audit trail a hand-coded scheduler never produces. Safety teams evaluating humanoid AI deployments should start demanding those transition logs in vendor testing — the record is already generated; almost nobody asks for it.
What buyers should ask vendors
Until transition learning ships commercially, buyers of humanoid AI hardware should ask one blunt question: how does your robot know a walking phase has ended, and what did that cost in demonstrations to teach? If the answer involves a hand-written scheduler and a five-figure demo budget, the seam will still be soft.
What Comes Next for Humanoid AI
From sequential to simultaneous
Professor Oh’s stated next step is to leave sequential behavior behind: not walking, stopping, then grabbing — but manipulating objects while moving. Whole-body, concurrent locomotion-manipulation is the tier where real work lives (think carrying a tray through a crowded room), and it raises every difficulty the RISE framework handled one at a time.
Beyond two behavior families
The responsibility-based expert training approach generalizes on paper: picking, bimanual manipulation, tool use — each is another specialty the scoreboard could recruit. The open question is how many specialists one humanoid AI can coordinate before the meta-controller’s weights become too noisy to read transitions from.
The safety conversation arrives with the skills
Transition competence is also a safety milestone: a robot that switches behaviors appropriately is one that can stop appropriately. Expect safety-simulation work — like the industry’s push to convince people that robot AI won’t hurt them — to adopt learned-transition systems as new test subjects, because hand-written schedulers at least fail legibly.
Three things to watch through 2027
First: independent replications of the simulation margins on third-party hardware. Second: whether open humanoid toolchains absorb responsibility-weighted experts within the year. Third: the first commercial deployment advertised as trained without combined task demonstrations — the tell that RISE-style economics reached procurement documents.
The research lineage to follow
RISE sits on a fast-moving lineage. NVIDIA’s GR00T family keeps pushing general humanoid foundation models; Figure’s control-stack work keeps landing in the humanoid AI conversation; and world-model research such as Google DeepMind’s Genie 4 world model teases training inside learned simulations. The papers worth watching combine cheap data, learned transitions, and whole-body dynamics — three constraints RISE satisfies two and a half of. If that intersection keeps producing, the debate moves from “does it work” to “what does it cost” by late 2027.
The scoreboard, visually
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
