Toward Self-Aware Robots

Abstract

Despite major progress in Robotics and AI, robots are still basically "zombies" repeatedly achieving actions and tasks without understanding what they are doing. Deep-Learning AI programs classify tremendous amounts of data without grasping the meaning of their inputs or outputs. We still lack a genuine theory of the underlying principles and methods that would enable robots to understand their environment, to be cognizant of what they do, to take appropriate and timely initiatives, to learn from their own experience and to show that they know that they have learned and how. The rationale of this paper is that the understanding of its environment by an agent (the agent itself and its effects on the environment included) requires its self-awareness, which actually is itself emerging as a result of this understanding and the distinction that the agent is capable to make between its own mind-body and its environment. The paper develops along five issues: agent perception and interaction with the environment; learning actions; agent interaction with other agents-specifically humans; decision-making; and the cognitive architecture integrating these capacities.

Keywords: Markovian processes; affordance; cognitive architecture; decision-making; human-robot interaction; learning; planning; self-awareness.

Figure 1

Architecture of the proposed sensorimotor approach for scene affordance learning.

Figure 2

Results from the perception process. Appearance and spatial information from the RGB-D point cloud of the real scene (Left); supervoxels from over-segmentation of the point cloud (Middle); and results from intrinsic clustering (Right).

Figure 3

Representation of the grasp-ability affordance relation.

Figure 4

Global action selection architecture composed of two decision systems implementing corresponding behaviors: the goal-directed expert is a model-based RL algorithm whereas the habitual expert is a model-free RL algorithm. The meta-controller is in charge of monitoring different expert information, giving control to one of the two. The reward information comes from the motivational system and represents the goal of the task.

Figure 5

(Left) Setup for the Human-Robot Interaction (HRI) task from Renaudo et al. (2015a): the human and the robot collaborate to put all boxes in a trashbin. (Right) Arena for the navigation task. A mapping of the states produced by the robot has been manually added. The red area indicates the goal location whereas the green areas indicate starting locations of the robot. Red numbers are starting location indexes; blue numbers are some states indexes referred to later.

Figure 6

Weights of each action (direct image of Q-values) for the habitual expert when alone (Top) or combined with the goal-directed expert (Bottom) at the end of the navigation task. Each light green dot is the final learned value of each action. The red bar indicates the best action to take from the human perspective. These measures are shown in the states next to the goal (s27, s28, s29).

Figure 7

Evolution of monitored signals when both experts are controlling the robot during the navigation task. (Top row) shows the sliding mean cost spent by both experts for decision-making. (Middle row) shows the measures of learning scaled by their coefficient. (Bottom row) shows the evolution of the probability of selection of each expert. In these experiments, the strong parameter of the habitual expert learning measure combined with its slow convergence favors the goal-directed selection in order to reach the goal more easily (however at a high computational cost).

Figure 8

Task of making an object accessible by the human to the robot (Pandey et al., 2013): (a) Places on the support planes where the human can put something with least effort. (b) Weighted points where the robot can support the human by taking the object. (c) The planner found a possible placement of the object on the box from where it is feasible for the robot to take. Note that, because of the object-closeness based weight assignment, this placement also reduces the human's effort to carry the object.

Figure 9

Initial state of the world in the Clean the table scenario. In this task, the robot and the human share the goal of cleaning the table together.

Figure 10

Shared plan computed by the robot to solve the joint goal: first removing the three objects (books) that are located on the table, then sweeping the table in order to clean it and finally placing the objects back on the table. While cooperatively achieving the task, the robot will be able to detect and assess correctly why the human partner stays idle, for instance in cases where, due to a momentary absence, the human may have missed the fact that robot has swept the table.

Figure 11

Decision-making architecture including operational, intentional and deliberation modules. The deliberation module implements a meta-reasoning capability.

Figure 12

The global cognitive architecture employed in the RoboErgoSum project. Blue modules are responsible for generating and managing the symbolic knowledge. Decision-making modules are shown in green. Solid and dashed lines are used only to improve diagram readability where lines cross, and are otherwise identical in meaning.