Robotics Briefing /
Carrying with the whole body, predicting trouble, and respecting preferences.
Moving an object is only part of useful robot behavior. A robot also needs to keep it supported, recognize an interaction heading toward failure, and follow instructions about how the job should be done. Today’s papers examine those problems, with evidence ranging from human-guided physical trials to a simulation benchmark for personal preferences.
Your quick takeaways
- Whole-Body Dexterity completed 45 of 50 physical trials using the arms and torso to support objects. A skilled human steered every trial; the full shelf-to-table sequence succeeded seven times out of ten.
- MAIA scores the likely contact problems in candidate movements before execution. Adding it to an existing policy raised reported physical task success from 52.6% to 61.5%.
- LIBERO-PeRM separates getting a task done from doing it the requested way. Strong task scores can coexist with poor preference satisfaction, even in simulation.
Papers
Whole-Body Dexterity: carrying can involve more than the hands
Conor Mc Gartoll and colleagues, Amazon, UC San Diego and University of Michigan • October 15 preprint
People often brace a box against their chest or tuck it under an arm. Whole-Body Dexterity gives a humanoid a learned controller for those changing contacts, using its arms and torso to support and reposition an object.
The system generates stable grasp configurations in simulation, trains specialist policies to move between them, and combines their behavior into one controller. On a physical Unitree G1, the controller refines a human operator’s motion commands to preserve support while the robot changes its grasp.
Across shoulder-touching, underarm-stowing and shelf-to-table tasks, it completed 45 of 50 trials. Directly mapping the operator’s movements to the robot completed 9 of 50, while the SONIC general whole-body controller completed 2. The total includes short repositioning round trips as well as the longer carrying sequence. On that full sequence alone, Whole-Body Dexterity succeeded in seven of ten attempts.
These are assisted-control results. One skilled operator, with more than eight hours of practice on the tasks, guided every trial and adjusted the controller’s guidance setting during the carrying sequence. Physical tests used a basketball and a small cardboard box, and occasional lost contact still caused drops. The evidence supports better control of body-object contact under human guidance; autonomous warehouse carrying remains a separate challenge.
MAIA: compare the contact risks before choosing a movement
Yi Qiao and colleagues, Shanghai Jiao Tong University • October 15 preprint
A motion can look reasonable while pushing a part into a jam. MAIA combines camera observations, planned actions and wrist force measurements to predict how an interaction will develop. A learned world model estimates future visual and force responses, and an anomaly predictor flags likely deviations.
The researchers collected 2,852 physical trajectories across eighteen tasks, including insertion, wiping, pressing, constrained movement and articulated objects. They annotated when abnormal interactions began, allowing evaluation of warnings before those events.
For robot execution, an existing π0.5 policy proposes eight action sequences. MAIA chooses the candidate with the lowest predicted anomaly probability, without retraining the underlying manipulation policy. Across eighteen tasks with thirty trials per task for each method, reported overall success increased from 52.6% to 61.5%. Wiping improved from 46.7% to 66.7%, while insertion stayed at 63.3%.
This makes monitoring useful for action selection, although it does not solve recovery after failure. Evaluation used one Franka Research 3 and a fixed force-sensor setup. New tasks required adapting the monitoring model, and the authors identify limited prediction horizons and missing cross-robot validation as constraints.
LIBERO-PeRM: completing the chore can still ignore the request
Zhixu Li and colleagues, UC Riverside, Georgia Tech and collaborators • October 15 arXiv release • NeurIPS 2026 Evaluations and Datasets
Putting a mug down may satisfy the basic task while ignoring a request to leave its handle facing outward. LIBERO-PeRM tests this distinction through 312 simulated tasks, covering movement preferences, skill choices and user-profile conditions.
The strongest tested model without preference-specific training, π0.5, completed 100% of the skill-preference tasks but achieved only a 25-out-of-100 preference-satisfaction score. That score allows partial credit and counts satisfaction only when the task succeeds; it is not simply the percentage of people who would be happy with the result.
Preference-focused training helped substantially in familiar layouts. However, its satisfaction score fell from 82.48 to 28.36 when the same tasks and preferences appeared in unseen layouts. Mixing ordinary task data with preference data preserved more general task ability, but weakened some movement preferences.
The benchmark provides a way to ask whether a robot followed the whole request. Its scope remains predefined, measurable instructions in simulation. It does not test evolving personal habits, long-term interaction with people, or whether the robot recognizes a conflict and asks a clarifying question. The project displays results and simulated examples; public code and data availability remains unverified.
News
Agility names Nyobolt as Digit 5’s battery partner
October 15 announcement • Update to the previously announced Digit 5
Agility identified Nyobolt as the supplier behind Digit 5’s planned fast-charging battery system. The ninety-minute runtime and nine-minute recharge specifications were already part of the robot’s September announcement; the new development is the named technology partnership.
The companies say the battery could support more than twenty hours of work per day, with early access expected in the first half of 2027. For warehouse operators, reducing charging downtime can improve how much work a fleet gets through.
Those operating-hour figures remain company expectations. The announcement does not provide a sustained customer-site trial establishing productive throughput, intervention rates or battery aging under that schedule.
NEURA introduces a model combining senses and memory
October 14 announcement • Catch-up • Company claims
NEURA’s Neuron AI combines vision, touch, audio, robot state and language with memory. The company says the shared model supports five of its robot bodies, with customer training and fleet updates through Neuraverse.
The intended benefit is handling information a camera alone cannot capture, such as resistance during contact or a sound confirming that a mechanism engaged. Remembering earlier observations could also help when an object leaves view.
NEURA plans an eleven-minute laundry-handling demonstration at GTC Berlin on October 20–22. That is a scheduled demonstration, not a completed reliability study. The announcement provides no comparable task-success table or independent validation of the model’s claims.
Read this first
Watch Whole-Body Dexterity’s shelf-to-table demonstration alongside its results table. The changing contacts make the capability easy to see, and the operator description and seven-of-ten full-sequence result give the demonstration its necessary context.
Read the paper