Towards a learnt neural body schema for dexterous coordination of action in humanoid and industrial robots

Abstract

During any goal oriented behavior the dual processes of generation of dexterous actions and anticipation of the consequences of potential actions must seamlessly alternate. This article presents a unified neural framework for generation and forward simulation of goal directed actions and validates the architecture through diverse experiments on humanoid and industrial robots. The basic idea is that actions are consequences of an simulation process that animates the internal model of the body (namely the body schema), in the context of intended goals/constraints. Specific focus is on (a) Learning: how the internal model of the body can be acquired by any robotic embodiment and extended to coordinated tools; (b) Configurability: how diverse forward/inverse models of action can be ‘composed’ at runtime by coupling/decoupling different body (body $+$ tool) chains with task relevant goals and constraints represented as multi-referential force fields; and (c) Computational simplicity: how both the synthesis of motor commands to coordinate highly redundant systems and the ensuing forward simulations are realized through well-posed computations without kinematic inversions. The performance of the neural architecture is demonstrated through a range of motor tasks on a 53-DoFs robot iCub and two industrial robots performing real world assembly with emphasis on dexterity, accuracy, speed, obstacle avoidance, multiple task-specific constraints, task-based configurability. Putting into context other ideas in motor control like the Equilibrium Point Hypothesis, Optimal Control, Active Inference and emerging studies from neuroscience, the relevance of the proposed framework is also discussed.

Appendix: A neural implementation of time base generator

A time base generator (TBG) is a scalar dynamical system in the normalized variable $\xi$ given by:

$$\begin{aligned} \dot{\xi }= & {} \gamma (\xi (1-\xi ))^{\beta }\nonumber \\&\beta \in (0,1), \end{aligned}$$

(5)

where $\xi (\hbox {t})$ is a smooth sigmoid from $\xi (0) =0$ to $\xi (t_f ) =1$, with a bell shaped velocity profile and desired finite movement duration $t_f$. The system has two equilibrium points, an unstable one at $\xi =0$ and a stable one at $\xi =1$, consequently the system always approaches stably to $\xi =1$. The time history of the TBG can be regulated using $\beta$. The $\gamma$ parameter has a dual function: controlling the convergence time and to reset the TBG and make it excitable for subsequent activation cycles. As regards to the exponent $\beta$, it can be shown that the condition,² $\beta >2/3$ is essential in order for the third derivative of $\xi \hbox {(t)}$ (Jerk) to be defined at $t=0$ and $t= t_f$. Under these conditions, it can be seen that the dynamics of the system are Non Lipscitzian,³ i.e. equilibrium configurations do not satisfy Lipschitz condition for ODE since $|\partial \dot{\xi }/\partial \xi |\rightarrow \infty$. This implies that equilibrium point is a terminal attractor, and systems with terminal attractor dynamics always converge in finite time (Zak 1991).

To derive the convergence time, let us consider a simpler dynamical system:

$$\begin{aligned} \dot{\xi }= & {} \gamma \xi ^{\beta }.\\ t_f= & {} \int _0^{tf} {dt=\int _0^1 {\partial \xi /\gamma \xi ^{\beta }=1/\gamma (1-\beta )} } \end{aligned}$$

Once again we can see that equilibrium point is a terminal attractor as convergence time is always finite and can be precisely specified through the constant $\gamma =1/t_f (1-\beta )$.

Remarkably, the above dynamical system can be approximated using a reciprocal inhibition network consisting of two neurons. A single neural element is an integrate-and fire neuron comprised of a multiplier, an integrator and a power function. In the integrate-and-fire model, input spikes are multiplied by their respective synaptic weights, summed and integrated over time. If the integral exceeds a threshold, the neuron fires and the integration restarts. The functionality in this case can be expressed as:

$$\begin{aligned} \dot{\xi }_i =\prod w_i \zeta _i \end{aligned}$$

where

$$\begin{aligned} {\zeta _i }=\xi _i ^{\beta } \end{aligned}$$

The reciprocal inhibition network of two neurons modeling the TBG is shown in Fig. 12.

Fig. 12

Reciprocal inhibition neural network for TBG

Dynamic behavior of the neuron can be written as

$$\begin{aligned}&\dot{\xi }_{1}=-\gamma {\xi }_{1}^{\beta }{\xi }_2^{\beta }=-\dot{\xi }_{2}\\&{\xi }_{1}(t)+ {\xi }_2 (t) =1\\&\therefore \dot{\xi }_{2}=\gamma {\xi }_1 ^{\beta }{\xi }_2 ^{\beta }=\gamma {\xi }_2 ^{\beta }(1-\xi _2 )^{\beta }=\gamma (\xi _2 (1-\xi _2 ))^{\beta } \end{aligned}$$

This is same as Eq. 5.

To perform any reaching movement, several joints—shoulder, elbow, wrist, fingers move cooperatively forming a synergy in a flexible and dynamic fashion. While groups of fingers may operate synergistically while playing a guitar chord, individual fingers are controlled while playing a lead. One of the basic problems of motor control is to understand how neural control structures quickly and flexibly organize and engage different parts of the body schema to cooperate synergistically in a movement sequence. The above TBG can be used to dynamically couple and decouple synergies in different ways based on task specification. In sum, by selecting two parameters of the TBG ($t_f$ and $\beta$), a family of time-varying signals can be generated. From the point of view of real-time implementation, it is possible to use any scalar function of time satisfying the properties of described above or a look-up table etc.