Super psyched for this competition!
While folding origami is not "economically valuable task" which many commercial labs care about outright, I love this challenge b/c:
1. Origami, at the highest level, is IMO one of the most complicated dexterous task human hands are capable of.
2. I reckon even for fairly simple origami designs, you'd need more than 2 fingers. So this is a task that really showcase the use of 5-fingered hands.
3. Folding origami involves lots of occlusion. In fact, I've seen an origami master fold a crane more or less completely based on touch. This means tactile and even force sensing would be required for complicated origami designs.
4. Origami is an art form with long cultural history in Japan. It's not just the physical manipulation techniques involved, there's also a creative aspect that I think will be a frontier goal for Embodied AI to shoot for.
5. We're co-organizing this challenge with the Nippon Origami Association, who are basically the human experts. This means we'll have human expert benchmark against what embodied AI can do.
I reckon this is such a difficult task that this competition will last for years before embodied AI (both model and hardware improvements) can catch up to human performance.
I'm also very confident that on the path of solving robotic origami we'd have also unlocked some fundamental manipulation capabilities that'd be useful in lots of other real-world dexterous tasks.
Grateful to be able to do this with my buddies @DJiafei@chris_j_paxton as well as strong partners from @SharpaRobotics@LightwheelAI@BitRobotNetwork as well as friends from academia.
I am excited to help organization of Origami competition in IROS 2027. We are releasing the certain amount of episodes for the origami airplane and the simulator to train and evaluate the policy.
If you are interested in this competition, please join us!!
Best robotic dancing yet! At this point most of general public would have seen Unitree G1s by now, even if they dunno it by the model name. Robotic entertainment is gonna go even more mainstream going forward.
https://t.co/xJNFNv8Qos
The bottleneck to on-robot reinforcement learning is good, scalable reward prediction. Robometer is a massive step in that direction, and the authors have been wonderfully open as well, releasing a large dataset and continuing to improve their model post release.
Thanks to @aliangdw@yigitkkorkmaz and @Jesse_Y_Zhang for joining me and @DJiafei!
Learning robust, general-purpose reward functions for robotics unlocks many potential applications, like on-robot reinforcement learning or dataset validation. However, thereβs a question of how to actually train such reward functions. Training success/failure prediction leads to ambiguous signals partway through a demonstration β itβs hard to measure progress β making the method unsuitable for reinforcement learning, among other things. Predicting progress, on the other hand, does not give a good way of using failure data.
So why not do both? Robometer combines both progress and preference supervision, resulting in a stable, scalable, and highly general reward learning approach. @aliangdw@yigitkkorkmaz
and @Jesse_Y_Zhang join us to tell us more.
Watch Episode #84 of RoboPapers, with Chris Paxton and Jiafei Duan today!
The Robotic Origami Challenge is coming to IROS 2026 (Sept 26) and it's one of the most creative dexterous manipulation benchmarks we've seen.
The task: train a policy to fold a traditional Japanese paper airplane.
The judge: an Origami Grand Master from the Nippon Origami Association.
You got #SharpaWave robot hands? Good!
You don't? We will be providing real-world eval support remotely.
Excited to see what you will come up with, along with the organizers @chris_j_paxton@micoolcho@DJiafei
More info: https://t.co/QubvM5dmaA