This paper discusses cognitive adequacy for robots collaborating with and assisting humans. We share insights from the development of robot architectures that use knowledge-driven and data-driven methods to jointly address challenges in transparent knowledge representation, reasoning, and learning in robotics. Consider a robot delivering objects to particular places or stacking objects in desired configurations. Such robots have to reason with different descriptions of incomplete domain knowledge and uncertainty. These descriptions include commonsense knowledge, e.g., relations between some domain objects and default statements such as “textbooks are usually in the library” that hold true in all but a few exceptional circumstances. At the same time, information extracted from noisy sensor inputs is often associated with quantitative measures of uncertainty, e.g., “I am 90% certain I saw the robotics book in the office”. Also, any robot in a practical domain will have to revise its existing knowledge over time, often using data-driven methods. In addition, for effective collaboration with humans, the robot may need to describe (or justify) its decisions and beliefs. In state of the art architectures that combine knowledge-based reasoning (e.g., for planning) and data-driven learning (e.g., for object recognition) for such integrated robot systems, cognitive adequacy, i.e., the ability to support the desired behavior, thus poses open problems in knowledge representation, reasoning, and learning. This paper builds on expertise in designing robot architectures [1, 2, 3] to identify some key underlying principles for cognitive adequacy.
Figure 1(left) is an overview of the architecture that encodes the principle of stepwise iterative refinement . It is based on tightly-coupled transition diagrams at different resolutions, and may be viewed as a logician, statistician, and an explorer working together. These diagrams are described using an action language, which has a sorted signature with statics, (Boolean, non-Boolean) lfuents and actions, and supports (deterministic, non-deterministic) causal laws, state constraints,
Coarser−resolution
Representation (Resolution 1) (Logician) Representation (Explorer) ianbtsetnrtaicotnal (Resolution i)
Interactive transition Learning observed outcomes
Representation (Statistician) (Resolution i+1) Finer−resolution
Representation (Resolution N)
Commonsense knowledge + theories of cognition, learning
Non−monotonic Logical reasoning Probabilistic Execution Probabilistic models of uncertainty Labels (training phase)
Features extraction Deincdisuiocntiotnree Current state
New axioms
Answer set Classification block
ASP program
Real scenes
Baxter Plan
Goal Answer set, domain knowledge
Text/Audio processing Raliextlieeorvmaalnss,t Ptreoxctessed
Program analyzer Outputs:
Output labels (occlusion, stability)
Explanations
(relational description)
and executability conditions. The domain’s history includes observations, action executions, and
prioritized defaults. For any given task, the robot plans and executes actions at two resolutions,
but constructs on-demand relational descriptions of decisions and beliefs at other resolutions.
Knowledge representation and reasoning: With two resolutions, the robot represents and
reasons with commonsense domain knowledge, including cognitive theories, in the coarse-resolution.
For example, a robot fetching objects in an ofcfie building reasons about the knowledge it has
about some attributes and default room locations of objects. It also has a adaptive theory of
intentions encoding principles of non-procrastination and persistence. The fine-resolution transition
diagram is defined as a refinement of the coarse-resolution transition diagram, introducing a theory
of observations that models the robot’s ability to sense the values of domain fluents. A robot in an
office building would (for example) now consider grid cells in rooms and object parts, attributes
that were previously abstracted away by the designer. The definition of refinement guarantees
that for any given coarse-resolution transition, there exists a path in the fine-resolution diagram
between states that are refinements of the coarse-resolution states. Also, the refined diagram is
randomized to model non-determinism in action outcomes. For any given goal, non-monotonic
logical reasoning at the coarse-resolution provides a plan of intentional abstract actions; this is
achieved using Answer Set Prolog, a declarative programming paradigm [
The architecture supports transparent reasoning and learning, i.e., explainable agency, by
encoding a theory of explanations comprising: (i) claims about representing, reasoning with, and
learning knowledge to support relational descriptions of decisions and beliefs; (ii) a
characterization of explanations based on representational abstraction, and explanation specificity and
verbosity; and (iii) a methodology for constructing explanations. This theory is implemented
in conjunction with the components summarized above—see Figure 1(right). The robot then
provides on-demand relational descriptions of decisions and beliefs in response to different types
of questions (e.g., descriptive, contrastive, counterfactual) posed by a human. The human is able
to interactively obtain descriptions at the desired abstraction, specificity, and verbosity, with the
robot posing disambiguation questions to the human as needed [
The following execution traces demonstrate the working of the architecture.
Execution Example 1. [Planning and Learning] Consider a robot in a that is asked to fetch a cup.
• The plan of abstract actions: (1, ℎ), (1, ), (1, ), (1, ), is based on the default knowledge that cups are usually in the ℎ. • For each abstract transition, the relevant (zoomed) fine-resolution description is identified, e.g., only cells in the and the ℎ are relevant to the first , and used to obtain a probabilistic policy that is invoked repeatedly to execute a sequence of concrete actions, e.g., robot is in a cell in the ℎ after first . • Suppose the robot’s attempt to pick up a cup in the kitchen fails. Using the knowledge that the cup is ℎ and its its arm is ℎ, the robot learns the axiom: impossible (1, ) if (1, ℎ), _ℎ(, ℎ), • When asked to provide a detailed description after plan execution, the robot revises the abstraction level to use the fine-resolution description.
Human: “Please describe the executed plan in detail.” Robot: “I moved to cell 2 in the ℎ. I picked the large cup by its handle from the counter [...] I moved to cell 4 of the . I put the cup down on the red table.” Execution Example 2. [Explanation and Disambiguation] In the simulated scenario in Figure 2, the human asks the robot to “Move the yellow object on the green cube.”. The reference to yellow object is ambiguous, and the robot asks for clarification. • Robot: “Should I move yellow duck on the green cube?”
Human: “No. Move yellow cylinder on the green cube.” • The robot computes a plan: pick up green mug; put green mug on table; pick up red cube; put red cube on table; pick up yellow cube; put yellow cube on table; pick up yellow cylinder; put yellow cylinder on green cube. • The robot traces beliefs and axioms to answer questions after plan execution.
Human: “Why did you not pick up red cube at step1?”
Robot: “Because the red cube was below the green mug.”
Human: “Why did you move yellow cube to the table?”
Robot: “I had to put the yellow cylinder on top of the green cube. The green cube was
below the yellow cube.”
Summary: Implementing principles such as stepwise iterative
refinement and relevance, and exploiting the interplay between
representation, reasoning, and learning, are key steps towards achieving cognitive
adequacy in architectures for robots. Such an architecture that
combines knowledge-based reasoning and data-driven learning has
providing promising results in simulation and on physical robots [
This work is the result of research threads pursued in collaboration with Ben Meadows, Tiago Mota, Heather Riley, Rocio Gomez, Michael Gelfond, Shiqi Zhang, and Jeremy Wyatt. This work was supported in part by the U.S. ONR Awards N00014-13-1-0766, N00014-17-1-2434 and N00014-20-1-2390, AOARD award FA2386-16-1-4071, and U.K. EPSRC award EP/S032487/1.