IROS 2026 will fill five days in Pittsburgh with 20 keynote talks, three plenary panels, three community forums and 86 workshops and tutorials. The 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems runs from 27 September to 1 October. Its programme also lists competitions, technical tours and a debate.
Programme structure
Sunday 27 September is given over to workshops and tutorials, with 44 planned for that day. A further 42 are scheduled for Thursday 1 October, so the sessions bracket the main conference days.
The keynotes take place from 28 to 30 September. Three plenary panels sit alongside them. The Monday panel covers start-ups and entrepreneurship, and the Tuesday panel covers space. On Wednesday the panel takes up a direct question: should robots be generalists or specialists?
Three forums, described as community discussion sessions open to all attendees, are held on Monday 28 and Tuesday 29 September. Monday afternoon is devoted to US government funding. Tuesday morning covers robotics education, with a focus on experiential learning. Tuesday afternoon asks whether robotics is underinvesting in its own foundations.
The programme also includes a science communication crash course at 11:00 on Tuesday 29 September. It is led by Sabine Hauert of the University of Bristol, Ella Scallan and Evan Ackerman of IEEE Spectrum.
The keynote line-up
Two of the 20 keynote titles are still to be confirmed, those of Gaurav Bansal and Nancy Pollard. The remaining 18 cover a wide range of subjects.
Surgery and social settings are both represented. Elena De Momi speaks on open questions on the road to autonomous surgery. Naomi Fitter addresses socially informed context awareness for robots in human-populated settings.
Several talks deal with AI and data. Genta Suzuki presents the Fujitsu Kozuchi Physical OS as a research strategy for scalable physical AI. Yekai Sun speaks on tokens of the physical world and accelerating intelligence emergence through multimodal embodied data. He Wang's title is Towards the AlphaGo and ChatGPT Moments of Embodied AI. Jen Jen Chung's talk is titled All data are equal, but some data are more equal than others. David Held covers robots that learn after deployment, with a focus on policies designed for fast adaptation.
Hardware and mechanical design get their own slots. Greg Chirikjian moves from hyper-redundancy to robot imagination and affordance learning. Yigit Menguc speaks on soft robotics and tensegrity. Gina Olson takes a biomimetic approach to snake robots, with the word soft flagged by a question mark in her title. Ralph Hollis covers Lorentz levitation technology for fine motion robotics, haptic interaction and vibration isolation. Lael Odhner asks where the frontier of manipulation systems sits today.
Industry and field practice feature as well. Al Rizzi speaks on developing advanced robot capability at the RAI Institute. Genki Yamauchi talks about extending construction robotics from big sites to the work done by hand. Tim Bretl addresses composite manufacturing. Robin Murphy rethinks disaster robotics and the open questions that matter now. Brittany Duncan draws on patterns from the field to show how contextual lessons can inform robot design. Ankur Mehta rounds out the list with robotics problems that robotics cannot solve.
What the schedule shows
Read as a whole, the schedule gives physical AI and embodied data more room than any other single theme. Three keynotes address it directly, and two more, on post-deployment learning and on data quality, bear on the same problem. The generalist-versus-specialist panel on Wednesday puts the same tension before the full audience.
The forums point to a second concern, which is how the field is funded and sustained. One session covers US government funding. Another asks whether robotics is underinvesting in its own foundations. A third covers experiential learning in education. Together they put resources and training on the agenda next to technical results.
Analysis: what the programme tells us and what it cannot
The mix at IROS 2026 matters because it places several questions in one venue that are often argued separately. Learning-based approaches to physical AI, mechanical design, field deployment and funding policy each have talks, and the panel on generalists and specialists links them. A roboticist working on soft hardware and one working on large embodied datasets will be in the same rooms on the same days.
Two features stand out from the titles. First, practice-oriented talks sit alongside the AI-focused ones. Murphy on disaster robotics, Duncan on lessons from the field, and Yamauchi on construction work done by hand all start from deployment conditions rather than from models. Second, several titles are framed as open questions. De Momi, Murphy, Odhner and Mehta each use that kind of framing. That suggests the organisers wanted talks that name unsolved problems, not only finished results.
Several things remain unproven, and the programme cannot settle them. A title such as Towards the AlphaGo and ChatGPT Moments of Embodied AI states an ambition. It does not show that the field has reached such a point, and a talk title is not evidence of results. The same applies to scalable physical AI and to fast adaptation after deployment. Those are research directions, and the schedule gives no benchmarks or outcomes for any of them.
The programme also leaves gaps in what can be said about scale. It lists no attendance figures, and two keynote titles are not yet confirmed. Comparing IROS 2026 with other conferences on size or influence would therefore go beyond the facts available. What can be said is narrower. The organisers have put funding, education and foundational investment in front of all attendees at the same level as the technical programme. For a field where much of the attention goes to demonstrations, that is a notable editorial choice, and the discussions in those forums are where it will show whether the choice was worthwhile.