Reflections from ICRA'26

First drafted: 2026 June 05
Last updated: 2026 June 13

A week has passed since I returned from Vienna and I'm still exhausted.

A week that felt like a month, given the breadth of topics discussed, the range of ideas shared, and the number of hands shaken. By the end, my always fleeting social battery was running on fumes, and it will take yet a couple more weeks to fully recharge.

This year, as opposed to what I wrote a few years ago, ICRA felt different though.

It may have been because I'm no longer a solo researcher grinding away at a niche topic from a lonely desk, brainstorming daily with the voice in my head. As a junior professor, my area of influence has now expanded, spreading across more research lines, multiples projects, and, at last, a team.

It may have been because, since becoming a father, I've grown far more selective about where I spend my time away from home. If I'm trading the comfort of my house and the chuckles of my 10-month old for a few days at a conference, it'd better be worth it.

Regardless, I found myself returning from Vienna with pages of notes, dozens of bookmarked papers, and far more ideas that I could reasonably process during the conference itself, keen to unload on the team back home.

Among those notes, a few observations stood out. I wanted to write them down here partly to share them with you, partly to pin this moment in time, to check back in a few years (or a few weeks, given the current pace of the robotics ecosystem) and see how much things have changed.

Many of my observations come from a space robotics perspective, but what struck me most is how much several research communities are beginning to converge around a common set of challenges.

Here are a few of the things I noticed at ICRA 2026.

Conversations shifting from models to data

One of the most interesting discussions I had was with Melanie Wille from QUT, who works on underwater robotics (where many of the challenges are quite similar to those we encounter in space). And it wasn't about new architectures or models, but about data.

Across several groups, I noticed a growing interest in understanding how much performance is actually determined by what we train on. How does image quality, environmental conditions, illumination and visibility, acquisition geometry, domain characteristics, or data composition affect downstream performance? Which properties of data matter most?

This may sound less glamorous than proposing a new neural network. But it cuts to the core of a fundamental problem for robotic systems aimed at operating in extreme and barely accessible environments: our systems rarely fail because they have too few layers. They fail because reality often differs from the conditions under which they were trained on.

Understanding domain effects may ultimately prove more valuable than endlessly scaling models.

And on that note…

Bigger models may not always be better

This idea was raised by Hiro Ono during his keynote at the Perceptual Challenges for Planetary Exploration workshop.

Navigating on Mars is hard, but from a visual perspective, Mars is actually much simpler than many terrestrial environments. Mars contains fewer dynamic objects (no one will run in front of the rover, that we know of), fewer semantic categories (sand, bedrock, boulders, steep slopes), and sometimes less raw environmental complexity than many Earth-based settings (say, autonomous driving in Jakarta).

Hiro raised an interesting question: instead of scaling models up, should we be investigating how far we can scale them down while maintaining mission-level performance?

Progress may not always require scaling up.

Sometimes it may require understanding which complexity is actually necessary for a given environment. For robotic systems operating under strict resource constraints, this distinction is crucial.

Nature as the best engineer for efficiency

Many of the talks I intentionally attended focused on operating under severe limitations: low power, harsh lighting conditions, limited communications, and scarce onboard computing resources. Autonomous spacecraft and rovers cannot rely on cloud computing, unlimited power, or constant human supervision.

Several talks, such as TJ's in the From Sea to Space workshop, explored neuromorphic sensing, event-based cameras, spiking neural networks, and biologically inspired computing. The common motivation was straightforward: how do we achieve reliable perception and decision-making under severe power and computational constraints? And can nature, honeybees for instance, give us the blueprint we need?

What I found particularly interesting was that many researchers are no longer treating efficiency as a post-hoc optimization problem. They are treating it as a primary design requirement.

That seems to me like the right instinct. The future of space robotics will not be solely determined by what systems can do under lab conditions or in ground-based demos under arguable assumptions, but by what they can actually do when operating on a few watts of power, in a place where no humans can intervene.

3D reconstruction as a foundational capability

I also noticed that across multiple domains, 3D dense reconstruction methods, based on gaussian splatting, neural rendering, and their (what seems like endless) variants, are rapidly finding their way into different pipelines.

Whether underwater or underground, many seem to be in search of richer, more accurate scene representations.

What makes this development exciting is that these richer geometric and photometric representations are increasingly becoming a foundational capability that enables navigation, inspection, planning, interaction and many other downstream tasks.

Space robotics does no longer feel a niche

Perhaps my strongest impression from ICRA 2026 was the growing maturity of the space robotics community. Across workshops, technical sessions, and hallway conversations, there was clear momentum around autonomous spacecraft operations, planetary perception, onboard AI, and long-duration autonomy. More importantly, the community feels increasingly connected to broader robotics research rather than isolated from it.

Challenges once considered unique to space are now relevant across multiple domains, while advances from terrestrial robotics are rapidly finding their way into space applications.

The flow of ideas is becoming bidirectional. And this is a healthy sign for the field.

To me, the most interesting developments in robotics are increasingly happening at the intersection of AI, physical constraints, and real-world deployment.

Whether we are building robots for Mars, the Moon, underground tunnels, the deep ocean, or industrial facilities, many of the underlying challenges are surprisingly similar. We need systems that can perceive under uncertainty, operate with limited resources, adapt to changing conditions, and make reliable decisions when human intervention is difficult or impossible.

Extreme environments at the edges of the spectrum of operational complexity expose the weaknesses of our algorithms and our hardware faster than any controlled laboratory setting ever will.

And judging by the conversations at this year's ICRA, the community is only beginning to explore what that future might actually look like.

D.R-M

Note: These opinions are my own. They are also highly dynamic and their second time derivative, albeit decreasing with age, is large. Experience provides me with new insights that drift former convictions. By the time you read this, assume my opinion on the subject has most likely changed.