Robotics Briefing /
Better predictions, better recovery, and the work around the robot.
A robot needs more than a plausible next move. It needs to anticipate what contact will do, recover when a task goes wrong, and fit into a system that keeps work moving. Today’s research tackles those problems from three directions, while new industry announcements put more attention on the software and support behind deployment.
Your quick takeaways
- RISE lets a humanoid choose how to stand up around obstacles. Small physical tests are promising, but following a requested path can still conflict with keeping enough support to rise.
- AW-0 combines reusable skill plans with predicted video and actions. It improved physical task scores, while changes in simulated appearance still caused a steep performance drop.
- Physics calibration helped a model plan under unfamiliar conditions in simulation. Separately, NVIDIA and Agility announced more infrastructure for training, deploying and supporting robots.
Papers
RISE: getting up without turning into the obstacle
Osher Azulay, Cheng-Lin Hsieh and Stella X. Yu, University of Michigan • October 21 preprint
A fallen humanoid may have several ways to stand. The nearest wall or raised platform changes which one is useful. RISE learns a repertoire of recovery movements, then uses depth images and a predictive planner to steer the intended motion toward a requested direction or route. Feedback control turns that plan into joint movements.
On a physical Unitree G1, unguided recovery succeeded in all 20 flat-ground trials and nine of ten staggered-step trials. Guided recovery succeeded in four of five trials from each of two flat-ground starting postures and all five trials near a wall. Those small denominators matter: five successful attempts do not establish dependable operation around people.
The paper’s wall comparison shows why perception matters during recovery, not just during normal walking. A blind version turns into the wall; visual feedback helps it change direction, and guidance favors the clear side earlier.
Inference ran on an external RTX 4090. Uneven terrain remains difficult because a desired path can conflict with the contacts needed to stand. Moving obstacles and external disturbances remain future work; code is listed as coming soon.
AW-0: connecting a reusable plan to predicted physical consequences
Zhuoyang Liu, Dantong Niu and colleagues • October 21 preprint
A robot can select a sensible sequence of skills and still execute the wrong movement. AW-0 joins code-based skill planning with a model that predicts future video and robot actions. The aim is to connect instructions and reusable routines to how the scene will actually change.
Across eight physical tasks on two bimanual robot platforms, AW-0 averaged a 71.1% task-rubric score, versus 46.0% for the strongest comparison model. This measures credited task progress, not an all-or-nothing completion rate. Each task used 100 demonstrations for post-training and 48 evaluation rollouts across three training seeds.
The harder result is in simulation: its success fell from 80.8% in clean scenes to 20.6% after visual randomization. A hybrid mode that also executes generated code improved the randomized result to 29.2%, but reduced clean-scene success to 63.1%.
Planning and prediction can complement each other, yet choosing between them introduces another reliability problem. Generated code can contain incorrect arguments, and the system does not learn continually from its own deployment. The evidence supports improved bounded tasks, not general autonomous work.
PhysiCalWorld: observe an interaction before predicting the next one
Yunhao Luo, Wentao He and colleagues, University of Michigan • October 21 preprint
Two objects can look identical and respond differently to the same push because their weight is distributed differently. PhysiCalWorld gives a predictive world model a short interaction clip from the current scene. It compresses that experience into a reusable description of the local dynamics, without fine-tuning the model for that scene.
Across four simulated environments, the researchers varied properties including mass distribution, joint damping and friction. In the Push-T task, adding the method to DINO-WM raised planning success from 53.3% to 62.7%. With a second model, LeWM, success rose from 48.0% to 61.3%. Each environment had 75 planning targets across 15 unseen physics configurations.
This suggests that a little evidence about how an object behaves can improve predictions beyond what appearance reveals. The test is simulation-only, however. Training requires trajectories grouped by their physical configuration, and the method assumes those properties stay constant during an episode. How well the interaction clip reveals the relevant dynamics also matters.
News
NVIDIA expands the path from robot training to deployment
October 21 announcement • Available software and planned hardware
NVIDIA announced Isaac Lab 3.0 as generally available and introduced the Isaac OS runtime, which connects models, sensors and control software. Agile Robots also announced the Agile ONE SN research humanoid, combining its body with NVIDIA’s compute and robot-learning tools.
The distinction is availability: the reference humanoid is scheduled for early 2027, and its Halos safety integration is planned. NVIDIA says its SimReady library offers more than 2,000 assets. The complete Human 1M dataset is restricted to Cosmos Coalition partners, so that dataset should not be described as fully public.
The practical promise is less integration work between training and a physical robot. The announcement does not establish reliable performance on a customer’s complete task, or certify every system assembled from these components.
Agility adds capacity for the work that follows a robot sale
October 21 announcement • Pittsburgh facility planned
Agility plans a 50,000-square-foot Pittsburgh facility for software, skill development, customer deployment and service. It expects employees to begin accessing the building by year-end. The site will include space to validate customer workflows and handle complex repairs, with an East Coast service center supporting deployed fleets.
This is a distinct development from the AGCO pilot covered yesterday. Its significance is the surrounding operation: testing a workflow, monitoring a fleet, repairing hardware and getting improved skills back into service all take people and infrastructure.
Agility also reports more than $300 million in multi-year Digit 5 orders, subject to contractual milestones. That figure is not delivered revenue or proof of completed installations. The expansion announcement supplies no measured improvement in uptime or repair turnaround.
Highlights
KINO-2 shows a crate-handling workflow, with faster model execution
Bear Robotics • October 21 launch • Company demonstration and internal measurements
Bear’s KINO-2 demonstration connects a warehouse order to a KR1 humanoid picking bottles, filling a crate, delivering it and returning with an empty one. The learning platform combines demonstration-based manipulation with the company’s navigation and workflow software.
Bear reports reducing GR00T model inference from roughly 500 to 60 milliseconds using NVIDIA’s runtime, TensorRT and quantization. That is a faster model calculation, not an eightfold increase in completed work. The company describes active manipulation pilots, but provides no task-completion rate or intervention count. Its large installed service-robot fleet should not be mistaken for thousands of deployed KR1 humanoids.
Read this first
Watch RISE’s wall comparison alongside its physical-results table. It makes a useful distinction visible: standing up, choosing a clear route and maintaining the support needed to finish are related problems, with different failure modes.
Read the paper