Robot learning, dexterity and a robotic thumb | AI robotics
Three new preprints provide physical tests for learning from video, safer robotic-hand manipulation and an open-source extra robotic thumb.
- robot learning
- dexterous manipulation
- robotic hand
- wearable robotics
- open source robotics
- AI
- news
What actually changed in the last 24 hours? Three preprints arrived with physical evidence worth inspecting: a method that learns from human and robot video, a hand that measures manipulation failures more carefully, and an extra robotic thumb with an open design. These are author-reported research results. They are not products or independent validation.
ThinkWorld uses human video to help learn robot actions
ThinkWorld, published on October 27, tries to use human video without labeled robot actions. The method represents motion as a sequence of tokens, predicts a future state, then reinforces sequences whose prediction matches what actually happened.
On its project page, the authors report 93.3% success across 120 physical trials and a 4,276-hour video training corpus. They also state that the reasoning component is used during training, not deployment. That matters: it does not show that any home camera can teach any arm a task.
The smaller-build lesson is simpler. When recording demonstrations, do not only keep the final gesture. Keep what changed before and after each action. That sequence helps test whether a policy learned the relationship between moving, touching and getting an outcome.
DexSafe measures dexterity alongside drops
The DexSafe preprint uses a LEAP hand trained in simulation to place objects within the hand. It adds a safety filter that can become more or less conservative when the object may drop.
The authors evaluate 4,800 unseen goals over 13.3 hours with four objects and several wrist orientations. Their project page reports 95.2% goals reached with the state-based policy and a 75% reduction in drops when the filter is active. The code is listed as coming soon, so this is not yet a recipe you can download and repeat.
There is a practical rule here. Do not score a gripper only by whether it finishes the task. Record how often it drops the part, how long it takes to detect trouble and how much success it loses when a safety limit is enabled. Those measurements say more than a video where everything works.
EgoThumb releases an extra thumb for physical tasks
EgoThumb, accepted at IROS 2026 and published on October 27, presents a robotic thumb that straps onto a hand. It puts the actuators on the forearm and uses a palm-facing camera to control two tendons from what it sees.
The project page links to the hardware and software repositories. The authors report four tasks, including unscrewing a cap and stirring tea, with 90% average success on the demonstrator’s hand. For two new users, they report 87.5% to 92.5%. The tests are still limited and author-reported, but there is open material here to inspect how the device is put together.
For robot builders, the value is not copying a whole thumb tomorrow. It is seeing a design that moves weight and motors away from the hand, keeps the light part near the fingers, and separates perception from action. You can start much smaller: one gripper, one fixed camera and one task you can repeat fifty times.
Every claim above comes from here. If something has no source, it's not in.