Better robot training meets the limits of the real world.

A Balanced Data Diet: more useful practice for simulated robots

Octi Zhang and colleagues, University of Washington and NVIDIA • October 8 • CoRL 2026 paper

A million simulated robots are not automatically a million useful learners. Many may repeat easy configurations or fail at ones they cannot yet solve. Success-Guided Sampling, or SGS, tracks recent success and emphasizes configurations near the robot’s current ability, while retaining access to the broader training distribution.

The study scales to more than a million parallel environments and tests both quadruped locomotion and assembly. It then transfers assembly skills to a physical UR5e arm by training a camera-based controller to imitate the simulated expert. No real-world fine-tuning is used for those deployments.

On hardware, gear insertion succeeded in forty-seven of fifty trials, rod insertion in thirty of forty-nine, and nut-and-bolt assembly in eighteen of forty-eight. First-attempt success was lower: 76%, about 29%, and 12.5%, respectively. Reported overall success therefore includes recovery attempts.

The contribution is making training experience more productive, but the physical results show how uneven transfer remains. Objects were positioned on a secured task board; shifting the receiving parts caused failures. SGS also selects from an existing set of configurations rather than discovering which situations the simulator should contain. The project has demonstrations, with code still marked coming soon.

• Paper ↗
• Project and demonstrations ↗

SimVLA: synthetic experience for a robot that moves and manipulates

Kyoungin Baik and Youngwoon Lee, Yonsei and Seoul National University • October 8 • Preprint

Moving a robot’s base changes what its cameras see and what its arms can reach. SimVLA addresses that coupled problem with synthetic robot actions across thirty-five tasks, visual question-answer pairs derived from simulator state, and additional simulated deployment trajectories.

A vision-language-action model connects images and instructions to robot movement. Here, its robot training data come from simulation, avoiding teleoperated task demonstrations. The evaluation uses an Anubis mobile manipulator in a mock kitchen, pantry and real home, including relocating objects, opening a drawer and pouring.

Across the evaluated settings, the paper reports 54.4% average task progress, outperforming its comparison trained on fifty real demonstrations per task. Progress is the share of predefined subgoals completed; a partly finished task can earn credit. It must not be read as a 54.4% full-task success rate.

The study totals 2,120 physical evaluation rollouts across policies and conditions. Coverage remains concentrated on rigid objects in kitchen-like settings, with only two task types tested in the home. It supports synthetic experience as a useful training source, not a claim of general household competence.

The public repository contains scene and data-generation tooling, but explicitly excludes the policy training recipe and checkpoint. That limits how completely others can reproduce the reported robot behavior.

• Paper ↗
• Project ↗
• Available code and release scope ↗

HuMBLE: humanlike walking that still follows steering commands

Mike Zhang and colleagues, RAI Institute, Boston Dynamics and Carnegie Mellon • October 7 • Preprint

Copying human motion can produce natural-looking walking but leave a robot unable to respond well to unfamiliar speed or direction commands. HuMBLE first learns from recorded human locomotion, then uses reinforcement learning to expand the commands it can follow while preserving characteristics of the reference gait.

The researchers demonstrate the approach on Atlas R1, Atlas D1 and Unitree G1. The resulting controller uses the robot’s internal state and a steering command to generate coordinated movement, rather than requiring a motion clip at each step. On G1, average policy inference took 1.1 milliseconds on the onboard Jetson’s CPU.

Physical demonstrations include changing walking direction and transitioning between slower and faster gaits. This is evidence of a compact, responsive locomotion component; the inference time is not a measure of the entire perception-and-control system.

The paper limits its core results to flat-terrain locomotion. Stairs, climbing and vaulting need additional motion references and terrain sensing. Demonstrations that combine HuMBLE with other controllers should not be mistaken for a single walking policy mastering every movement.

• Paper ↗

Sereact robots enter Zalando returns operations

October 7 • Joint company announcement of live operations

Sereact, Zalando and CEVA report that dual-arm robots are processing fashion returns at facilities in Greven, Germany, and Świebodzin, Poland. Their Cortex system identifies, grasps and sorts returned items without separate training for every product, according to the partners.

Returns are a demanding test because garments arrive in inconsistent shapes and packaging. The deployment puts that variability inside a working logistics operation, with staff responsible for supervision, exceptions and quality control.

The announcement identifies real sites but does not disclose robot counts, intervention rates, throughput or uptime. Those missing measurements matter when judging how much work the systems reliably handle.

• Joint announcement ↗

Odyssey-3 opens a research preview

October 8 • Company release following its September introduction

Odyssey’s new preview lets users interact with generated environments and see how they respond to actions. The company also reports robot-arm adaptation using tens of hours of demonstrations, including recovery from missed grasps, and humanoid-policy work with Flexion.

A world model predicts how a scene changes. If those predictions capture the right physical relationships, they may help train or control robots across different settings.

The release reports video-prediction and generated-world benchmark results, but those scores do not establish reliable physical manipulation. Its robotics demonstrations remain company-reported, without a detailed robot trial table in the announcement. Treat the preview as a way to inspect the model, rather than evidence of a ready-to-deploy general robot controller.

• Announcement ↗

OpenViTac gives touch a more explicit test

Yifan Wu and colleagues, Fudan-led team • October 7 • Benchmark preprint

OpenViTac pairs simulated and physical tasks involving texture, fragile objects, contact and precise insertion. Its OpenVTLA model reached a 54.6% unweighted average success rate across eight physical task categories, versus 41.0% for the vision-based π0.5 comparison.

Each task uses a separately trained policy, and most physical comparisons contain only twenty trials. Touch helps expose information cameras miss, but sensor calibration and the gap between simulated and real tactile signals remain constraints.

• Paper ↗
• Benchmark overview ↗