Context: In this article we match machine learning (ML) and interactive machine learning (iML) with radical constructivism (RC) to build a tentative radical constructivist framework for iML; we then present a pilot study in which RC-framed iML is applied to assistive robotics, namely upper-limb prosthetics (myocontrol. Problem: Despite more than 40 years of academic research, myocontrol is still unsolved, with rejection rates of up to 75. This is mainly due to its unreliability - the inability to correctly predict the patient’s intent in daily life. Method: We propose a description of the typical problems posed by ML-based myocontrol through the lingo of RC, highlighting the advantages of such a modelisation. We abstract some aspects of RC and project them onto the concepts of ML, to make it evolve into the concept of RC-framed iML. Results: Such a projection leads to the design and development of a myocontrol system based upon RC-framed iML, used to foster the co-adaptation of human and prosthesis. The iML-based myocontrol system is then compared to a traditional ML-based one in a pilot study involving human participants in a goal-reaching task mimicking the control of a prosthetic hand and wrist. Implications: We argue that the usage of RC-framed iML in myocontrol could be of great help to the community of assistive robotics, and that the constructivist perspective can lead to principled design of the system itself, as well as of the training/calibration/co-adaptation procedure. Constructivist content: Ernst von Glasersfeld’s RC is the leading principle pushing for the usage of RC-framed iML; it also provides guidelines for the design of the system, the human/machine interface, the experiments and the experimental setups.

Key words: Machine learning, interactive machine learning, radical constructivism, assistive robotics, human-machine interaction, co-adaptation.

Upshot: Our commentators’ very constructive criticisms point out a number of weaknesses in the design of our experiment, and offer insight into how such weaknesses might have led to the poor results of the experiments. We summarize the suggestions, which point in a few precise directions, and outline how we will try to implement them in the near future.