TacSushi keeps failed food-handling attempts as training lessons ↗
The robot learns from recorded consequences, but unfamiliar ingredients remain difficult. The work appears in this week’s CoRL workshop program.
Embodied AI
ORBIT AI lab revises experiments with human checkpoints ↗
ORBIT connects AI planning with physical experiments, while people approve revised plans.
Iontronic robot skin handles pencil lead and triggers protective reflexes ↗
A layered readout tackles the traffic from thousands of tactile sensors.
CEL robots share successful practice to improve phone packaging ↗
A 96-robot training fleet helped reduce test failures from 25 to two, while people still labelled outcomes and reset tasks.
DexSeed expands 19 human grasp sequences into robot training data ↗
Generated demonstrations support a robot-grasping research system.
Machenta robots learn a new task from one demonstration ↗
The company demonstrates rapid adaptation across robots; deployment reliability remains a separate question.
A robot chose the right box using its partner’s view, while sharing 80% fewer visual tokens ↗
Astribot’s CoMind research transmitted compact, task-specific visual features between two real robots. A robot picked the correct obscured target in 16 of 18 trials, using one-fifth as many tokens as a generic visual-sharing baseline.
AI wrote its own robot flipping tricks, reaching 95% in simulation. Real objects were harder. ↗
Berkeley researchers let a coding agent revise a robot controller after simulated failures. Its best tuned controller succeeded in 342 of 360 simulated trials, but physical tests on three unfamiliar objects succeeded only six times out of nine.
An AI practiced, took notes, then tackled 14 household tasks a robot policy struggled with ↗
A coding agent carried its frozen practice notes into simulated household tests. It completed more goal conditions on selected tasks, but each run took about 38 minutes and cost tens of dollars.
Teach a robot by moving its arms: Flash ALOHA opens its training code ↗
The release connects demonstration collection, policy training and robot control. Researchers still need their own hardware, recordings and trained models.
A robot's footage helped locate a forgotten jacket. AgiBot opens its continuous-observation tool. ↗
LIFE searches an organized record of a robot's day. Real-time code added November 10 lets developers connect their own cameras, while the full memory experiment remains an offline replay.
Adding touch made a robot worse at zipping a bag. Translating the signals lifted its score from 15% to 59% ↗
Robbyant's LingBot-Touch converts different tactile sensors into shared features. In the team's robot tests, raw sensor data could hurt performance, while encoded touch improved it. The percentages award credit for task steps, rather than counting only complete successes.
Robot makers pay up to about $150 for an hour of training video. The people filming it can earn about $1.20: inside China's embodied-AI data supply chain ↗
An investigation by China's Sci-Tech Innovation Board Daily, published Nov. 11 and republished by 36Kr on Nov. 12, found that an "hour" of robot-training data can sell for anywhere from 5 yuan to over 1,000 yuan. After five or six layers of subcontracting, some of the people recording the footage earn as little as 8 yuan an hour, and the same clips are often copied and resold.
Wear a Tiny Camera, Film Your Chores, Earn $6 to $18 a Task: Memories.ai Pays Ordinary People to Teach Robots ↗
On Nov. 10, US startup Memories.ai launched Memories Upload, a gig platform that pays people to wear a sub-45-gram LUCI Pin camera and film first-person video of food prep, shelf stocking and repairs to train robots. Fortune reports $6 to $18 per task, and the company says beta users logged over 1 million hours in the first month.
NTT gives robots a physics world model that previews a move before they make it, lifting a lab stacking test from 73.1% to 92.3% ↗
In warehouses and airports, boxes come in all shapes and weights, and a robot's reach can collapse a stack. On November 11, Japan's NTT said it has built a physics-based "world model" that generates a short video of what would happen for each candidate action, then picks a move less likely to drop something. NTT says a small-robot shelf-loading test improved from 73.1% success to 92.3%. It plans field trials with Mujin on airport baggage. Those results are from NTT's own experiments.