In this conversation, we spoke with Christian Duriez, CEO of Compliance Robotics and Research Director at Inria’s DEFROST Team, about the constraints shaping soft robotics, why deformation itself becomes part of the control system, and what would need to change for compliant robots to reach industrial maturity.
How do you frame the control problem of a soft continuum robot when deformation itself becomes part of the system’s intelligence?
First, it means accepting that you can’t control everything. Controlling a deformable robot starts with acknowledging that complete control is impossible. Then, part of the “control” can be delegated to the mechanical system itself. For example, regarding precision: for a long time in robotics, we pursued excellent absolute accuracy, but in many tasks in unstructured environments it is more effective to have good relative accuracy—leaning on a table, picking up an object placed on it without knowing coordinates, adapting to a wide variety of shapes, etc. In such cases, the robot’s compliance intrinsically provides this relative precision.
In real-time FEM-based control, what constraint most shapes your engineering decisions today?
Choosing FEM-type models is not dogmatic. It is mainly driven by the need to account for the properties of deformable materials used to build our soft robots and to unify the model’s use from design to control, including a digital twin updated by sensors on the robot. That said, FEM introduces computational constraints, especially for real-time control. Model reduction methods are currently the most mature way to address this.
When multiple configurations can achieve the same end position, what principle guides how your system selects one over another in production?
We compute inverse models through optimization. When several configurations are possible, we add a term that minimizes deformation energy. The system therefore selects the configuration that minimizes energy consumption.
Hollow replaces traditional grasping with internal transport. What operational limitation were you trying to eliminate?
We aim to improve pick-and-place operations by guaranteeing both throughput and safety.
When contact is expected rather than avoided, how does that change how you think about motion planning?
It doesn’t change anything. Our new inverse solver computes motion by optimization and “natively” includes contact handling.
At what point did simulation stop being a validation layer and become part of the control architecture?
Once it became possible to compute the robot’s inverse model (inverse statics / inverse kinematics) via an optimization method in real time, we started to base our control architecture on simulation. We also followed a trend seen in rigid robotics, while adapting the models and solving strategies.
Soft systems promise intrinsic safety. How do you translate that into measurable industrial guarantees?
We will follow the same directives and standards as collaborative robots. To obtain industrial guarantees, these standards define admissible forces exerted on humans depending on the body parts in contact. As with robots on the market today, we can limit speed and especially applied forces, but installation safety also depends on the application. If a soft robot is integrated with a sharp tool at the end-effector, it remains dangerous. I don’t believe in the promise of fully intrinsic safety.
Over extended deployment, what material behavior most challenges model accuracy, and how do you maintain reliability?
Over extended deployment, the material behavior can vary strongly due to creep, fatigue, micro-cracks that can appear, and delamination when using composites. Depending on the case, the material can stiffen or soften (Mullins effect with elastomers, for example).
What signal convinced you that soft robotics was ready to move from research to an industrial product?
The key for an industrial product is that it has a market. I see many use cases where there is a market for soft robots because real needs exist. At Compliance Robotics, in particular, we focus on fragility (of manipulated objects, the environment, and the robots themselves), and there are many industrial needs around that.
As a founder coming from deep research, what assumption did you have to abandon when entering the market?
That customers will all have a GPU capable of computing a highly detailed FEM model in real time to control the robot. That is why model reduction methods appear more advantageous.
With SOFA open-source, where do you believe your durable advantage truly sits?
In a fast-moving field like robotics, I don’t believe there are truly “durable advantages” for any company, especially for a startup. However, researchers are sensitive to open-source because it enables knowledge sharing. For our customers, it is an assurance they won’t be “trapped” by closed software that could suddenly stop being maintained. Our approach is therefore to build user communities around our technology, and open-source around SOFA is a tremendous driver for that.
If deformable robotics reaches structural maturity, what would the first undeniable industrial proof look like to you?
There are several ways to provide proof. For me, I would define some KPIs around execution quality and speed, control of forces applied to the objects, the users and the environment, the resources and energy consumed during manufacturing and use of robots, and ease of integration.
