Exploring Efficient Data Management in Robotics Software with Antoine Bassoul

Antoine Bassoul - Robotics Software Engineer

1. Introduction to Antoine Bassoul and His Expertise

• Question 1: Antoine, could you please introduce yourself to our readers and tell us more about your background in robotics software development and operations supervision?

I’ve worked in autonomous mobile robotic software development for the past 10 years. For 6 of those years, I was at Continental, one of the world’s largest automotive suppliers with 200,000 employees, during the self-driving car boom. My focus was on developing Radar perception algorithms such as detection, tracking, and SLAM.  It was an incredibly exciting time—everything was new, fast-paced, and we had to figure out almost everything from scratch because the field was so fresh. Things like do we compute on the device or on a centralised unit, what physical interface, com protocol, and middleware do we use and of course all the perception algos were not as mature as they are today. Applications were also very diverse, ranging from small agile team kind of projects, for self driving car startups, prototypes such as localisation in a gold mine or for race cars to very large consumer car projects with hundreds of people involved, automotive grade compute units, hard real time and safety constraint, thorough and established validation process.

Then I moved back to France and took the lead of the perception team at Navya, a self driving buses startup. So different from Conti, but very impressive, in the 2010’s they had designed a bus from the ground up, made it autonomous and sold over 200 units around the world. So we were operating at a much smaller scale than Conti, but it’s a robotic operation that had already expanded quite a bit and we were definitely facing scaling issues. That’s where I became fascinated with improving the software workflow of robotic teams, realising how much friction there can be in robotic software development once a company starts to expand.

Most recently, I joined Exail as the lead of the autonomy team for Drix, an unmanned surface vessel (USV) designed for ocean exploration missions lasting up to 10 days. While remotely operated, Drix can autonomously execute missions and avoid dangers. Currently, there are around 30 units in operation worldwide

Exail USB inspector

• Question 2: What sparked your interest in robotics and how did your journey lead you to your current role?

In the 90s in France we had the french robotic cup (the “E=M6 cup” for the connoisseurs, which I believe was one of the first of its kind) airing on TV, and at the time I couldn’t think of something cooler than those robots, that’s probably where it all started. In high school I was programming industrial robots and video games and I hadn’t to think twice about studying electrical engineering. There I got really obsessed with mathematics and that got me to signal processing. I was really into programming and maths, but like many roboticists I think, I had to deal with something “physical” so tech wasn’t for me and I started my career designing Sonar and Radar data fusion algorithms for the French Navy. Then in the 2010’s self driving cars were such a hot, innovative and rapidly growing area, I couldn’t imagine doing something else. After gaining a bit of experience I managed to land a job in Germany and moved without any hesitation.

Equipments for experiments in a mine

2. The Importance of Data in Software Development and Operations Supervision

• Question 3: In your opinion, why is data so crucial in robotics software development and operations? How does data impact decision-making and overall performance in your field?

It's a vast topic, but data management is crucial for successfully scaling a mobile robotics business. A key aspect is testing and validation, which represents a significant portion of the development effort in robotics software. Much of this is done by replaying both real and simulated data, and increasing the number and diversity of scenarios replayed enhances the software's reliability

For example, let’s say you already have a few hundred self-driving buses operating for a few years across various customer sites, and with a fairly mature autonomy stack. A customer reports a new edge case that could be easily solved by tuning an existing algorithm. Due to the nature of robotic systems, even a small change to address an edge case could cause regressions in other scenarios, which is why it must be thoroughly tested by replaying and analyzing many hours of recorded data. The validation effort increases as the number of robots grows and uptime requirements become more demanding, making it essential to automate the process as much as possible. In the beginning, the development team might be able to handle, store, and replay some recordings locally, along with conducting field tests, but this quickly becomes far too time-consuming.

To give an idea of the effort required without proper data management and automation: imagine having 30 customer sites, each with one robot equipped with a lidar, and performing one hour of log replay to validate each site. This would result in over 1 TB of data stored locally on each development station, which would need to be replicated every time the database is updated with new curation, tags, indexes, annotations, or scenarios. Assuming the system is re-simulated in real time, this could mean 30 hours of replay, plus the same amount of analysis time as overhead for each iteration loop — even for small tuning adjustments. Roughly speaking, you would either need to add this overhead to each pull request, or space out software deliveries and perform a 'big bang' integration and validation at the end of each product increment.
Building tools to automate replay and analysis is most effective when done collaboratively by infrastructure, QA, simulation, and algorithm teams. This requires a well-maintained, accessible, and high-performance validation database. The database must be curated, tagged, indexed, and annotated incrementally as new edge cases or scenarios are added, with contributions from development teams, operations, and QA. I’m a bit obsessed with automated large scale testing for robotics. 

A second key point is monitoring and observability. Robotic software systems generate vast amounts of heterogeneous, multimodal data, so being able to efficiently interpret all this data to understand what’s happening inside a robot at any given time is crucial. This is important in real time for the development team, QA, or customer support to address development needs or handle emergencies, but also offline to efficiently triage and analyze issues

A third key point is analytics. For example, comparing the percentage of time spent in autonomous versus manual operation across the fleet between two software versions, or generating a heat map of where strong braking or manual takeover events occur on a given deployment site. This is essential for understanding how customers are using the robots and for improving the quality of service.

Lastly, there’s deep learning and artificial intelligence. Without good data management, even building a basic video detector would require creating a dataset, which is time-consuming and costly, adding friction for the development team driving the initiative. This greatly hinders how quickly and efficiently these techniques can be adopted. On the other hand, in a robotics company with excellent data management, it is easy and fun for an engineer to take initiative and build deep learning models for various applications, potentially speeding up development, simplifying maintenance, and improving performances. Furthermore, I believe there will be many opportunities in the near and mid-term to use artificial intelligence on a much larger scale in the robotics sector. These technologies will likely become key differentiators for companies that successfully implement them to generalize their software stack and expand into more use cases and embodiments. We could even reach a point in the near future where you don’t need to modify the software at all — simply showing or instructing the robot to behave differently will be enough. The way to prepare for this is by collecting more, and higher quality, data points.

Stay tuned for part two of our interview with Antoine Bassoul!

As we wrap up this first part of our conversation with Antoine Bassoul, we’ve explored his impressive background and the critical role data management plays in robotics software development. But there’s much more to come! In the next article, Antoine will dive deeper into the pressing challenges faced by robotics teams today, the innovative tools driving efficiency, and how companies like Heex Technologies are shaping the future of data management in the industry. Don’t miss Part Two!