Dear #AcademicTwitter:
This is a friendly reminder that you're people first and researchers second. Taking mental health days is completely legit.
The bar to take time is not an election/earthquake/court ruling... you are always entitled to protect your mental health.
Teleoperation makes dexterous manipulation harder than it needs to be. The operator gets no direct haptic feedback, and controlling a robot hand remotely is rarely intuitive.
At @origami_robo , we've been working on EXOperation: a wearable robotic hand mounted on a gravity-compensated arm. Operators guide the robot from within: demonstrate a task, pull your hand out, watch the same hardware replay or use the collected demonstrations to train a policy.
EXOperation enables intuitive, rapid, and highly replayable human demonstration for robot learning. Check out the video and figure below for a glimpse of our ongoing work.
Thanks to everyone who came by Booth 205 at #CoRL2026 to try S1 and our Gen2 hand!
Due to WIFI issues, some sign-ups from the booth iPad and QR code didn’t reach us. If you left your details please take a minute to send them again:
origami-robotics.pages.dev/c…
We’ll follow up about availability and pricing. You can also email founders@origami-robotics.com.
Sorry for the hassle, and thanks again for coming by!
Quanting made it clear to me on day 1 -- until we fundamentally rethink actuator design from the ground up, we simply cannot achieve dexterity.
He had a list of problems with every hand out there. Every road led back to the actuators.
So he built a team and ...
Launching project flow. Our new generation robot hand.
Direct drive, liquid cooling.
From now, the heating problem is solved for robot hands.
more soon… @origami_robo
We’re excited to announce our collaboration with @PalatialSim to build the Dexterity Factory, a real2sim2real pipeline designed to bring human dexterity to robot hands.
By combining Origami Robotics’ S1 skill-capture gloves and robot hands with Palatial’s simulation technology, we’re building a pipeline that connects real-world human demonstrations, simulation, and robotic manipulation.
This is a step toward scaling how robots learn to manipulate everyday objects.
And we’re just getting started.
Next week, we’ll release our first 100 open-source datasets to support the robotics research community.
The glove and robot hand share the same joint arrangement. After calibration, each measured angle gives us a target for the corresponding robot joint.
Watch the knuckle, middle and tip joints move independently. We want to preserve that detail in the recording.
We tried 30 demanding tasks while wearing S1, including fast Rubik’s Cube solves, card shuffling and butterfly-knife tricks.
The glove needs to let people perform the movements we want to record. These demonstrations show what they could do while wearing it.
A million hours of what?
If we’re scaling human demonstrations for robot learning, what should each hour contain?
We built S1, our skill capture glove, alongside our robot hand to record joint movements that map to robot targets, with touch and video on the same timeline.
I am at #COLM2026 this week in SF:
> presenting a paper on reasoning behaviors in language models with @Jeande_d, @ybisk, @gneubig on Thursday (more info in thread below)
> co-organizing the Workshop on Learning from Situated and Embodied Interaction on Friday 🤖 (learning-situated-interactio…)
Excited to hear about everyone's research and happy to chat about work that I have been doing on multimodal social intelligence in AI agents across embodiments!
🧵New Paper: Reasoning models now think out loud for thousands of tokens, often correcting themselves, testing hypotheses, and hedging. But are these behaviors actually associated with correct answers? Largely, no.
Looking at accuracy, these models are doing great and often outperform their instruct counterparts. But accuracy does not tell the whole story. It does not tell us which behaviors showed up in the trace, which failures happened along the way, or whether the behaviors thinking amplifies are the ones most associated with correctness.
In our new paper, we study whether thinking models amplify the reasoning behaviors most associated with correct answers. This is important because a behavior can become more frequent without being a sign that the answer is right. For a quick example, a model may branch over several hypotheses without testing the right one, or revise its wording without fixing the reasoning underneath.
We introduce Behavioral Lift, which measures how much correctness differs when a behavior is present versus absent. Across 15,282 annotated traces from 15 models on 6 reasoning and knowledge benchmarks, we find a consistent mismatch. Thinking models strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment. But the highest-Lift behaviors change much less: confidence calibration, self-awareness, and knowledge alignment. We call this the Amplification-Lift Gap.
The attached figure shows the gap: no behavior is both strongly amplified and strongly predictive of task success.
ArXiv: arxiv.org/abs/2608.13760
Work with an amazing team: @Jeande_d, @lmathur_ , @ybisk, Robert Hawkins, @gneubig
Diffusion / flow-based robot policies unlock two axes of test-time scaling: sequential (denoising steps) and parallel (samples). Both improve performance but cost latency on a robot, and knowing which to scale a priori is often unclear. ELASTIC learns it via RL!