Latest / Embodied AI 101 / Bending Paper, Shaping Dexterity: The Robotic Origami Challenge

Transcript

Bending Paper, Shaping Dexterity: The Robotic Origami Challenge · Embodied AI 101

• Bending Paper, Shaping Dexterity: The Robotic Origami Challenge. Picture a grandmaster of origami moving inches of delicate paper between nimble fingers—not once, but six precise times—to create a perfect paper airplane. Now imagine replacing those fingers with robot hands and teaching them to perform the same task. That’s the new Robotic Origami Challenge unveiled at IROS 2026: a benchmark pitting dexterous robotic hands against Japanese origami experts. It sounds like a whimsical contest, but it tackles a serious problem in robotics: how to achieve human-level fine manipulation. Origami folding embodies many core challenges of manipulation — handling a deformable object that permanently changes shape, planning a complex sequence of moves, and coordinating two hands with surgical precision. As Devin Balkcom and Matt Mason observed over a decade ago, origami “is a fresh challenge for the field of robotic manipulation” that highlights “key problems” in how robots perceive, plan, and manipulate foldable structures. The new Robotic Origami Challenge builds on that vision by providing a rich benchmark: hundreds of real teleoperated fold demonstrations, simulated environments that faithfully model paper physics, and a scoring rubric decided by origami masters themselves. In this episode we’ll break down what the challenge entails, why it matters, and how researchers are already training policies to fold faster than the human pros. Origami and Robot Dexterity: A Hard Match. Why origami? On the surface, folding paper seems simple—people do it as children—but roboticists know the devil is in the details. A stiff block or rigid tool is one thing; paper is another. It starts flat and pliable but gains rigidity where you crease it. Every fold is a discrete, irreversible change: once you make a crease, the rest of the paper’s motion must respect that new geometry. The paper remembers its creases, permanently carrying that “fold memory” into subsequent steps. Even a small misalignment early on can foil the entire final shape. And holding a thin, slippery sheet of paper requires just the right compliance and force control—too soft and the paper crumples, too hard and you rip it. Past research has long noted these challenges. As Balkcom and Mason first pointed out in 2008, origami brings together classic manipulation issues in one package. They built the first origami-folding robot and a folding-sequence planner, noting how every crease adds complexity. Later work (for example Namiki and Yokosawa, 2015) applied dynamic motion primitives to folding, again stressing how difficult even “paper manipulation” is for multi-fingered hands. Those papers emphasized that, while humans make origami seem effortless, automation must solve perception, contact modeling, and long-horizon planning simultaneously. In other words, origami is hard but, if conquered, offers a benchmark to test a robot’s finest skills. In the last couple of years, the AI community has also started to use origami as a cognitive testbed. For example, Agarwal et al. (2026) introduced OrigamiBench, an interactive simulation where a model tries to propose folds and check if the resulting shape matches a goal. They found even large vision-language models struggle: scaling up model size “does not reliably produce causal reasoning about physical transformations” and the models “fail to generate coherent multi-step folding strategies”. In short, folding paper isn’t just about pixel classification or LLM prompts — it really tests a system’s ability to reason about cause and effect, geometry, and physics all at once. The IROS Robotic Origami Challenge brings this problem right into the robotics arena. It’s organized in partnership with the Nippon Origami Association (the masters of the craft) and robotics teams (LightwheelAI, SharpaRobotics, etc.), under the BitRobot grand challenges network. The stage is IROS 2026 in Pittsburgh: a real competition where each robot team must fold the same paper airplane model—called a kami hikōki, a traditional six-fold Japanese plane—under timed conditions. Human origami masters from the Nippon Association set the bar: they will judge whether a robot’s output is a recognizable airplane, and rank teams by how quickly they can complete the folds. In other words, not only does the robot have to nail the fold geometry (crease accuracy, symmetry, structural fidelity and no tears), but it must also do it faster than everyone else. The first prize can be up to $1 million — a clear statement that solving fine-motor skills is a grand AI goal. The Challenge Task: Folding a Flying Origami. Let’s be concrete about the task. Every competing team must fold an identical sheet of paper (a default 15×15 cm square, at least 60 gsm weight) into a classic paper airplane. The target is a “traditional Japanese paper airplane” folded in exactly six moves. In origami terms, that’s fairly simple classification of folds, but it still requires dexterity. The rules say the plane will be judged by an origami grandmaster: if the final model is a recognizably good airplane (“pass”), every successful attempt is scored by finishing time. Faster folds rank higher, and if a robot even manages to have the plane glide or fly, that’s extra bragging rights (“let the sky be the reward”). Go over 10 minutes and it’s an automatic fail, so no dawdling. The score is binary success/fail by appearance, plus timing. This means the robot’s policies should aim for both correctness and speed. Folding time becomes the tiebreaker. In practice, that sets a high bar: a skilled human can do this airplane fairly quickly, often around 20–30 seconds in casual practice. For reference, Grand Masters might do it a bit under 10 seconds (just folding the paper itself, not flying it). So robot teams target sub-human times. If a trained policy can consistently fold the model faster than an origami expert, it “wins” in spirit. According to the authors, their learned policies in simulation are indeed able to outperform the human baselines on fold time, showcasing that AI can match and even exceed the origami masters in this specific task setup. The difficulty is clear. The robot must sense and manipulate an initially flat sheet, execute a fold (which typically involves bending at an edge and pressing), then repeat five more folds. Each fold must be precise: a small mis-rotation means the airplane comes out off-center or unsymmetric (very un-Japanese). Moreover, once the sheet is folded, some parts are tucked or stacked, which constrains subsequent moves. The sequence is fixed and known (teams are given the diagram for the standard six-fold pattern), but the execution requires dexterous coordination of two hands. Most folds like this need one side to hold the paper taut while the other side pushes down a crease. This makes it a bimanual manipulation problem — a subject of rising interest in robotics. Bimanual tasks, by definition, require tight coordination between two arms. As Peng et al. (2026) point out in their new BiCoord benchmark, effective bimanual manipulation demands “continuous inter-arm dependency and dynamic role exchange across multiple sub-goals”. In other words, the hands must constantly know what the other is doing, often swapping active/passive roles (one holds, one moves). Origami embodies this: one hand might hold a corner while the other flips an edge, then they switch roles for the next fold. No surprise then that recent work on general bimanual benchmarks (e.g., BiCoord) highlights it as a hard, long-horizon problem. Folding a paper airplane is a specialized example of such long-horizon, tightly coordinated bimanual tasks. To succeed, a robot needs more than raw control: it needs subtle tactile or visual feedback, compliant finger surfaces, and robust planning. Traditional cameras or motion tracking can help it see the paper’s state, but the controller must adapt on the fly. Also, typical rigid-body simulators fare poorly here because paper is neither fully rigid nor fully soft cloth — it has a mix of elastic and plastic behaviors. The Physics has to allow sharp creases (plastic deformations) and then hold them (viscoelastic memory). We’ll see how the benchmark’s simulators tackle this challenge. Benchmark Datasets and Simulation. A key aspect of this new benchmark is that it provides everyone with the same training resources so that the only difference is one’s algorithms, not access to data or hardware. Organizers have prepared three main tools for every team: (1) a teleoperated demonstration dataset, (2) fold-sequence diagrams and example outputs, and (3) a high-fidelity simulation environment. These are released to each registered team (“barrier to entry is policy, not logistics” they say). Teleoperation Demonstrations (500+ episodes). First and foremost, the benchmark comes with over 500 human-executed folding demonstrations. These are full episodes of an expert operator controlling a robot (or pair of robot arms) to fold the six-step airplane from start to finish. In practice, that might’ve been done with a teleoperation rig: an expert origami folder uses, say, VR controllers or physical motion-capture gloves, which translate their hand motions directly to the robot hands in real time. The result is a large dataset of trajectories: joint angles or end-effector movements over time, coupled with sensor readings (like camera images or force sensing) and the known state of the paper mesh. Such a dataset is gold for imitation learning or offline reinforcement learning. With 500+ episodes, you have tens of thousands of state-action pairs covering every fold situation. The teleop data is accompanied by fold diagrams and example final folded shapes, so one knows exactly what the goal is at each sub-phase. In other words, the paper comes with the sequence description: at fold #1, align the corners; at fold #2, flip an edge, etc. Teams can use all this to bootstrap policy training. For example, they might first do behavior cloning on the human trajectories, giving their model a head start. (We’ve seen in other dexterous tasks that cloning a few hundred expert demos can teach a policy much of the structure of the motion.) Then one could refine the policy with reinforcement learning in sim, or use offline RL algorithms on the dataset. The point is, no team has to start from scratch or collect their own demos; the human expertise is already encoded in these 500+ episodes. Fold-Sequence Diagrams and References. In addition to raw demos, the organizers explicitly provide the “fold-sequence diagram” and “reference example folds”. A fold diagram is basically an origami crease pattern: it shows the two-dimensional layout of creases needed to go from the flat sheet to the final shape. The reference example folds might be images or 3D scans of the correct intermediate steps or final models as folded by an origami grandmaster. These clarifications help ensure every team knows exactly what the airplane should look like at each stage. In practice, a robot controller can use the fold diagram as high-level guidance: after fold 2, the paper’s shape should match reference image X. This tightly couples perception and action: the robot not only executes motions, but must possibly verify them visually (checking if the crease is sharp, if edges align, etc.). Those resources effectively remove ambiguity about the task: it’s not “fold any plane you want” but a specific target model with known geometry. High-Fidelity Simulation. The third pillar is the simulator. Teams get access to a virtual environment (based on NVIDIA Isaac Sim) that models the exact folding scenario. Crucially, this simulation is physically realistic for thin sheets: it treats the paper as a thin-shell that can crease plastically and remember its folds. Standard rigid-body physics won’t do: the developers needed to tune the paper’s mass, stiffness, crease strength, and friction to match real behavior. For instance, after one crease is made, the simulated paper will hold that bend even if forces are removed, just like real paper would. Moreover, the sim includes digital twin models of the partner robotic hands. Team sponsors (LightwheelAI, SharpaRobotics, etc.) each provided CAD and kinematic models of their hardware, so the sim can recreate their exact hand geometry and reach. This is big: it means a control policy trained in sim is working with the same kinematics as on the real robot, easing sim-to-real transfer. You won’t have discrepancies like the paper clipping through a virtual finger or the robot lacking a joint. In practice, a team can take the provided Isaac Sim scene, plug in a reinforcement learning library or trajectory optimizer, and train their folding policy end-to-end. They can also use domain randomization (varying paper friction or slight hand jitter) to make the policy robust. The simulation even comes with camera sensors placed as in the real robot, so vision algorithms can be developed there. If a team uses MuJoCo or PyBullet instead, sim accuracy might suffer, but Isaac’s goal is to be the canonical training ground. According to the challenge announcement, these assets will be accessible soon (“Coming soon” on the site), ensuring all teams have the same starting point. By providing data and sim, the challenge lowers the entry barrier: it becomes a question of algorithmic innovation. Who can best leverage the 500+ demos and the sim to produce a policy? In the closed world of the simulator, companies can iterate rapidly, then finally deploy on hardware to tweak any remaining reality gaps. This is reminiscent of how OpenAI tackled the Rubik’s Cube with Shadow Hand: lots of simulation and domain randomization, fed by policy gradients, allowed them to surpass human speeds on a tortured puzzle. Here, the same principle applies to paper folding. Training Policies for Origami. With data and sim on hand, how do teams typically train a folding policy? The paper itself (and the summary we have) suggests the teams explored combinations of imitation and reinforcement learning. Although we don’t have the exact technical details, we can extrapolate plausible approaches from prior work in dexterous manipulation. A straightforward pipeline is imitation learning: use the teleop dataset to train a neural policy (or behavior); essentially a supervised regression mapping observations (like camera images or hand senses) to actions (joint velocities or forces). Because the dataset contains expert human solutions, a cloned policy can often achieve a high success rate in simulation. The high-performance graphics simulator can run hundreds of episodes faster than wall-clock, so one could even do iterative DAgger-style imitation (where the policy tries something, sees where it fails, and those mistakes become new training points). However, pure imitation has limits. It might copy the easiest motions but fail to adapt when things go slightly off-course. That’s where reinforcement learning (RL) comes in. In sim, one can reward the robot for bringing the paper into the intermediate folded shapes (as defined by the reference diagrams) and for completing folds. Combining RL with the demos is a hybrid approach: perhaps initialize the policy via behavior cloning, then refine with RL. This is analogous to how RL has been used to fine-tune policies for in-hand manipulation or cloth folding. For example, Caldarelli et al. (2024) tackled cloth folding by using a clever model-based MPC on top of cloth physics models. They noted systems often fold clothes too slowly or imprecisely compared to humans, so they learned surrogate linear models (a Koopman-based approach) to speed things up. The origami problem is easier than full cloth (because the paper doesn’t flap like a blanket), but it similarly requires accurate yet fast control. An RL algorithm could learn, for instance, when to apply which force and how to rotate the sheet quickly without causing a misfold. Another idea, given the time-based scoring, is to explicitly train for speed. The policy could receive negative rewards for time taken, or a bonus for earlier completion. This encourages it not to linger unnecessarily. Indeed, the challenge rules explicitly count folding time, so the fastest fold wins. In practice, researchers might run trials where the reward includes “1/time” or “-time” once the fold succeeds. The summary hints that the trained policies did exactly that: they achieve faster folds than humans. It could be a side-effect of greedy RL or an objective. Importantly, the policy must generalize from the specific cases in the demo set. The curriculum is fixed (always the same six folds), but in deployment there could be small differences (like slight misplacements or slippage). Sim randomization helps: vary the paper mass slightly, or the friction a bit, so the policy doesn’t overfit. Vision can be integrated too: some teams might use camera input to predict the paper’s crease configuration. If a vision network can identify exactly the paper’s pose at each step, it could correct for any misalignment before the next fold. That could lead to closed-loop controllers, which adjust on the fly, rather than using a purely open-loop demonstration. On the hardware side, success depends on the gripper design. The digital twins in sim cover all the sponsored hands, but the hardware also needs to detect contact and maybe even pinch force. Some solutions might use compliance or fingertip rubber to get a firm grip on the thin paper edges. Others could use suction or magnets (unlikely here) to hold the paper. The policy might incorporate force sensing: e.g., fold until it feels the paper bend. One more twist: some papers on robot folding (like those by An et al., 2018) have even explored self-folding origami, where pieces of paper autonomously change shape. That’s different (micro-scale with heating, for example) and not directly relevant. Our scenario is full tele-operated or autonomous folding. It’s more akin to robotic cloth manipulation competitions (in previous IROS/ICRA events), but cloth retains fluid-like behavior and doesn’t have “crease memory.” Thus origami demands an additional modeling of the plastic hinge at each fold. Finally, there’s the baseline of pure teleoperation: one could imagine just building a refined teleop pipeline (like DexPilot style) where the robot is tele-operated in real time during the contest. But the idea of this benchmark is to train an autonomous policy. Indeed, the challenge description emphasizes “AI models” to control the robot, not just human-in-the-loop control. Teams will likely show off their autonomous folding algorithms, given the benchmark nature. Putting Robots to the Test. How do we know if the robots are doing well? First, each attempt is judged pass/fail by an origami master: this mimics how competitions like the Rubik’s cube or cloth folding competitions have human judges for quality. If the robot folds a shape that looks enough like the target airplane, it’s a pass; misalignment means fail. The master’s opinion is final. This addresses a key difficulty: automatically evaluating an origami shape is itself non-trivial (images vs models). By using human judgment, the benchmark bypasses building an automatic crease-validation algorithm. Once a robot’s attempt is successful (which is a tough bar!), the only metric that matters is time. The rules explicitly say “Faster successful folds rank higher.” So two teams that both fold correctly—in, say, 25 seconds each—would tie, but 20 seconds would beat 25. And if a plane actually flies when thrown (maybe the airport demonstration?), well, “let the sky be the reward” suggests they might give an informal extra applause, though it’s not formally scored beyond showmanship. In recent robotics challenges, humans and robots have been compared on speed or efficiency. For origami, the organizers actually benchmarked human performance via their own experts. The summary claims the trained AI policies outperform the human experts. In practical terms, this probably means that on average the robots folded faster than the humans did in demonstration. (For instance, if humans took around 30 seconds on average, perhaps the robot policies can do it in ~15–20 seconds consistently under sim.) That is both exciting and expected: once a robot has a good policy, it doesn’t get tired and can be perfectly consistent. Of course, outperformance is fragile in the real world: slipping a finger or slightly uneven lighting can cause a mistake. That’s why the challenge has a real-world finale: at IROS 2026, teams will bring real robots to do this live. The simulation is a training ground, but the robot must eventually operate on a physical paper. This is similar to Dexterity Challenges in other venues (Rubik’s Cube, block-stacking, etc.), where simulation helps but real trials decide the winner. By then, though, teams should have used the provided data to pre-train their models, and then fine-tuned on the actual hardware (maybe via on-robot RL or just extra calibration). Context and Related Work. This origami benchmark sits at the intersection of several research threads. On the “dataset and sim” side, it echoes endeavors like Robonet and RoboTurk for cloth and household tasks. But those often focus on vision; here the focus is on tactile-dexterity with deformables. The provision of 500+ teleop demos is reminiscent of other large teleoperation efforts. For example, Shaw et al.’s recent “Bimanual Dexterity for Complex Tasks” (CoRL 2024) collected human teleop data for grasping and kitchen-style tasks. DexPilot (ICRA 2020) is another example of collecting teleop data for dexterous hands via vision-based control. These projects show how useful real human demonstrations are. The Origami Challenge takes that to the next step by packaging it as a publicly shared benchmark for the first time in origami. In the domain of material manipulation, clothes are a common analog. Cloth folding tasks have been featured in competitions, and research like Caldarelli et al. (2024) has shown how planning for fast folds is hard in simulation-to-real transfer. They even needed a Koopman-based surrogate model to fold quickly under an MPC framework. Origami tasks share the sequential planning aspect but differ because the material becomes stiffer after each fold (cloth generally doesn’t do that). Still, lessons apply: a successful origami policy may also need to “plan ahead” — deciding on a strategy that anticipates future folds. On the other hand, classical robotics papers on origami (e.g. Tanaka et al. 2008, “Origami folding by a robotic hand”) have focused on solving geometry and kinematics for simple models. Such work often assumed perfect state knowledge and deterministic controllers. The new challenge modernizes that by incorporating uncertainty (vision, contact friction) and learning. Instead of hand-coded plans, we use data-driven control. Another interesting angle: since vision-language models are hot, one could ask if something like GPT-4 could plan these folds. Agarwal et al.’s OrigamiBench suggests current LLM-powered approaches aren’t yet reliable at sequential folding. That’s why we still need the kind of specialized benchmark and learning shown here. Eventually, perhaps a vision-LM could parse the fold diagram and direct a robot in words; the benchmark provides the ground truth to test such ideas. Lastly, this fits the theme of “general dexterous skills.” BitRobot, the grand-challenges organization behind this, has other missions like assembly (e.g. robotic IKEA), or navigation (Mars rovers). The origami challenge is one strand in a portfolio aimed at tackling robotic problems with direct human analogs. By putting a big prize and a structured benchmark behind it, they hope to accelerate progress — and it’s doing just that by releasing data and making comparisons fair. Outlook: Next Folds. What comes after this origami benchmark? If it succeeds in spurring new capabilities, we may see more folding-related tasks. For instance, a future challenge could be folding more complex models (multi-colored dragon, anyone?), or assembling origami 3D shapes. The fold mechanics learned here could transfer to things like folding packaging or flexible circuit boards in manufacturing. Even outside origami, the techniques for simulating and learning with plastic deformation could benefit other domains (like metal bending co-processes or plant modeling). From a research perspective, this benchmark invites work on planning under uncertainty, sim-to-real transfer, tactile sensing, and bimanual coordination. The fact that baseline policies can already beat humans in speed suggests there’s room for creativity: maybe policies that fold at superhuman speed but still with 100% success. Maybe we’ll see hybrid strategies where a coarse grasp-and-fold is refined by small corrective actions (wedges and re-grasps). The challenge might also encourage new hardware: perhaps specialized grippers or end effectors designed just for thin-paper folding. In summary, the Robotic Origami Challenge is more than a contest; it’s a shed in which the community can try out ideas on a well-defined problem. By providing the data and tools, the organizers let everyone focus on the algorithms. And by pitting robots against the wisdom of origami masters, the challenge ensures we’re pushing the envelope of dexterity. Already, reports indicate learned controllers can fold faster than those masters — a milestone in robotic skill. For the embodied AI community, the message is clear: start practicing your virtual folding now, because at IROS October 2026 those robots will need to be ready to fold (and maybe even glide) to victory. Sources: Prior work and key context include classic robotic origami studies, the new OrigamiBench AI evaluation, the current IROS challenge overview and rules, as well as related bimanual and folding benchmarks. These references highlight why paper folding is challenging, how the new benchmark is set up, and how it links to broader robotics goals.