A Bio-Inspired Mechanism for Learning Robot Motion From Mirrored Human Demonstrations

Abstract

Different learning modes and mechanisms allow faster and better acquisition of skills as widely studied in humans and many animals. Specific neurons, called mirror neurons, are activated in the same way whether an action is performed or simply observed. This suggests that observing others performing movements allows to reinforce our motor abilities. This implies the presence of a biological mechanism that allows creating models of others' movements and linking them to the self-model for achieving mirroring. Inspired by such ability, we propose to build a map of movements executed by a teaching agent and mirror the agent's state to the robot's configuration space. Hence, in this study, a neural network is proposed to integrate a motor cortex-like differential map transforming motor plans from task-space to joint-space motor commands and a static map correlating joint-spaces of the robot and a teaching agent. The differential map is developed based on spiking neural networks while the static map is built as a self-organizing map. The developed neural network allows the robot to mirror the actions performed by a human teaching agent to its own joint-space and the reaching skill is refined by the complementary examples provided. Hence, experiments are conducted to quantify the improvement achieved thanks to the proposed learning approach and control scheme.

Keywords: imitation learning; robotics; sensor-based control; spiking neural networks; visual servoing.

Figure 1

The robotic manipulator and human arm sharing the same end effector position and jointly moving during motor babbling to provide a proper training data.

Figure 2

Schematic diagrams for motor cortex-like map (MCM).

Figure 3

(A) The symmetric spike-timing-dependent plasticity (STDP) learning rule at various values for τ_a and τ_b. (B) A plot of change in synaptic weight vs. training iterations.

Figure 4

(A) A three link planar arm employed for numerical simulation. The robot deviates from a target path ϕ while moving from the current pose to the target pose to move through ρ instead. (B) The data collection schemes are illustrated. Motor babbling (in the upper panel) commands linear motion in joint space under the guidance of a differential forward kinematic DFK solver. Target path follower (in the lower panel) under the guidance of a differential inverse kinematic DIK solver.

Figure 5

Schematic diagrams for SOMs connected through Oja-Hebbian plastic synapses. This architecture allows correlating the joint spaces of the human arm and robot arm. During the training phase, BMUs (in AJ-SOM and RJ-SOM) from both maps that fire together are more likely to have an increase in strength of the connecting synapses. Consequently, during the control phase, if the same BMU in AJ-SOM becomes active, the corresponding node in RJ-SOM becomes active as well.

Figure 6

A schematic diagram of the correspondence of the human and robot joint spaces along with the task space TS. The data collected from the human and robot together allows building correlation (i.e., f(q_h) between JS_H and JS_R). This allows generating more examples to train the map g(q_r, v) correlating TS to JS_R by transforming examples conducted by the human arm (in TS) into JS_R.

Figure 7

A schematic diagram of the full system including both the static (SOMs network) and differential(MCM network) maps formed. Blocks with an inclined arrow passing through are the ones where learning/adaptation occurs during each phase.

Figure 8

(A) The human arm while moving in straight paths and (B) the plot of the Cartesian position of the hand-held object while moving.

Figure 9

Heatmaps of the standard Kohonen SOM depicting the relation between each input from the joint spaces of (A) the human arm and (B) robot agents.

Figure 10

Heatmaps of the varying density SOM depicting the relation between each input from the joint spaces of (A) the human arm and (B) robot agents.

Figure 11

(A) The robot arm while moving after training by the transformed data and (B) the plot of the Cartesian position of the end effector while moving.

Figure 12

Values of the fitness function vs. the number of optimization iterations.