IROS 2026: forget backflips. These problems decide whether robots become useful
Featured visual: RoboMorrow. Original editorial cover, not event photography.
The most important robot at a conference does not necessarily perform a backflip. It may simply handle an unfamiliar object more reliably, or let a researcher reproduce an experiment without weeks of manual configuration work. Those are the changes we are following in RoboMorrow’s IROS 2026 report.
IEEE/RSJ IROS takes place in Pittsburgh from 27 September to 1 October. This edition is a 28 September 2026 snapshot, based on the programme and published research rather than our attendance on site. An earlier preprint does not become a new announcement merely because its authors present it at this conference. [18]
What is confirmed, and what remains on the radar?
Georgia Tech’s programme includes AGILE, OG-VLA and KEYGEN. They offer three useful entry points: transferring skills from simulation to hardware, incorporating spatial information into robot action models, and generalising manipulation across objects. A programme entry alone does not confirm how a session unfolded or establish the reliability of a commercial product.
We distinguish a research paper, a demonstration and a deployment. A paper can document a controlled experiment. A video can show a single successful attempt. Deployment adds questions about repetition, exceptions and maintaining the system. The categories can support one another, but they are not interchangeable.
| Area | Source material | The practical question |
|---|---|---|
| Humanoid skill development | AGILE — learning and evaluation workflow | Can the result be reproduced on another configuration? |
| Geometry and language instructions | OG-VLA — orthographic representation | What happens when the viewpoint changes? |
| Generalisation across objects | KEYGEN — keypoint-based representations | Does the skill transfer to an unfamiliar object in the category? |
AGILE: less manual rescue of an experiment
AGILE proposes a structured workflow spanning environment verification, training, evaluation and deployment. The authors describe experiments on Unitree G1 and Booster T1. Here, AGILE is a research project, not Agile Robots or its Agile ONE humanoid. [19]
This addresses an important but relatively unglamorous problem. If an experiment works only on its author’s computer, it is difficult to build a product around it. A change of robot, task or environment can trigger another sequence of manual adjustments. A systematic evaluation process can expose regressions before they reach physical hardware.
For a non-specialist, the point is straightforward. We should not ask only whether a new skill looks convincing. We should ask whether the team has a repeatable method for discovering when it works and when it stops working. The second capability makes systematic progress easier to distinguish from a carefully prepared performance.
The paper does not imply that installing a package turns any G1 into an unattended industrial worker. Research results remain tied to stated tasks, hardware and experimental conditions. Moving a method into another process requires a separate assessment. A good evaluation workflow can support that assessment without replacing it.
OG-VLA: a robot needs geometry as well as language
OG-VLA combines vision-language-action capabilities with a spatial representation. It uses RGB-D observations and orthographic views to reduce sensitivity to camera pose. The first preprint appeared in June 2025; presentation in the IROS 2026 programme does not change its publication history. [20]
Consider the instruction “put the object in the box”. Understanding the sentence is not enough. The robot must locate the object and destination, choose an end-effector orientation and execute the motion in physical space. A small change in viewpoint can be a larger obstacle than the fluency of a language interface suggests.
For deployment, the interesting question is therefore not the result in one scene but robustness to a changed perspective. Does moving the camera require collecting a new dataset? Does the representation retain the information needed for an accurate grasp? What happens when an object is partly hidden? These questions keep language capability separate from physical reliability.
Percentages in research summaries also need context. An improvement over a particular baseline is not the same measurement as a task success rate. We have not promoted a relative improvement into a headline without the corresponding benchmark and measurement definition. A research contribution can be significant without pretending that its benchmark describes every real workplace.
KEYGEN: a different object should not require starting again
KEYGEN investigates object representations and category-level generalisation, using keypoints to describe objects. This is a different problem from merely recognising that an image contains a cup or a tool. [21]
Manipulation depends on relationships relevant to action. Objects can vary in size, material and shape while remaining in the same functional category. Studying such representations may reduce the need for a separately prepared behaviour for every individual object. That is a promising direction, not a claim of working on anything a person can pick up.
A meaningful practical test would include objects that were not used to tune the method. The set should contain more than nearly identical variants. Explicitly describing the differences is what allows a reader to understand how broad the reported generalisation actually is.
The distinction matters to a buyer because a flexible-looking demonstration can hide a narrow object catalogue. Expanding that catalogue might be easy, or it might be another research project. The published evaluation should help reveal which interpretation is justified.
Why every demonstration does not become a separate news story
Before publishing a standalone report, we need a primary source, an identified platform and a clear description of the demonstration. “It worked first time” deserves particular care. It could mean the first attempt after substantial preparation, not the absence of training or integration. Without the procedure, those interpretations cannot be separated.
Different hardware generations must also remain distinct. A Digit v4 video is not a test of Digit 5. The independently sourced story about the new generation’s design and availability is covered in our Digit 5 analysis. Identifying the platform is part of the fact, not a minor caption detail.
Five questions that turn conference enthusiasm into an assessment
First, did the experiment include unfamiliar objects and conditions, or only those used while refining the method? Second, how many attempts were made, including unsuccessful ones? Third, when and how did an operator intervene? Fourth, what happens after an error: automatic recovery, restart or manual repositioning? Fifth, which parts of the work are available for reproduction?
These questions are not intended to diminish research. They establish what has actually been achieved. A limited experiment can be valuable when it clearly demonstrates a new capability and its boundaries. The problem begins when a bounded result is presented as a finished solution for arbitrary work.
A prospective buyer should add an operational question: can the solution be maintained after deployment? A demonstration does not by itself explain updates, diagnostics, spare parts or responsibility for the complete process. Those are subsequent engineering requirements, not automatic consequences of a publication.
A useful internal review could therefore keep two columns: the capability demonstrated and the work remaining before procurement. That makes it possible to appreciate scientific progress without turning a research budget into an accidental production commitment.
Report status and subsequent updates
28 September 2026 — baseline edition: confirmed conference dates, three selected research directions, primary research links and criteria for evaluating demonstrations. No future-dated entries or unverified accounts of completed sessions have been added.
This URL is the reference point for subsequent verified updates. A changed date should represent a material addition, not an automatic refresh. Continue with our German Physical AI analysis and humanoid buying guide.
Sources and methodology
Editorial analysis based on the linked sources, checked on 28 September 2026. Manufacturer and seller statements are not RoboMorrow test results. We have not independently tested these robots.
Georgia Tech — IROS 2026 programme and dates · Zhao et al. — AGILE, arXiv:2603.20147 · Singh et al. — OG-VLA, arXiv:2506.01196 · KEYGEN — authors’ project page · Agility — Digit 5 announcement, 15.09.2026