A cognitive architecture (CA) implements a principled and cognitively inspired theory of robotic intelligence that provides capabilities for: 

• Maintaining common ground on task and human states (working/episodic memory)
• Reasoning about percepts and goals (inference)
• Planning and executing actions (problem-solving, skill retrieval)
• Learning from experience (skill learning)
• Integrates lower-level robotics capabilities
• Allows for high-level approaches to Task Planning and Behaviour Generation which,
  • Is Goal-driven rather than State-driven
  • Allows for abstract skill specification instead of hardware-specific programmes
• More intelligently handles exceptions
• Provides significant benefits like reusability and generality compared to ad-hoc robotic architectures

[Watch the video] The ability of the CA for maintaining object permanence (i.e., remembering/tracking objects that go out of visual view).

Collab AI developed a new method to teach the robot to recognise new objects through visual self-exploration. This method is called “Teaching with Active and Incremental Learning for Object Registration (TAILOR)”.

About TAILOR: 

• Provides an interactive process for training a deep-learning based object detector
• Alleviate the need for costly and tedious data collection and annotation
• Allow non-expert users to teach the system to recognise new objects without the need for programming

Collab AI developed a Collaborative Human-Robot Intelligent System (CHRIS) which allows a non-expert user to automatically program a robot through teaching demonstration.

• Teach and interact with robots in natural and efficient human-like manner
• Teach robots to learn task procedural knowledge from human demonstration
• Transfer learned skills to new tasks and robot execution with minimum prior knowledge

CHRIS was demonstrated at 2021 Hannover Fair Messe and was a Top 5 finalist for the Kuka Innovation Award 2021.

Collab AI developed a general key-point based end-to-end learning framework for skill learning for low-level tasks:

• Zero-shot sim-to-real transfer
• No camera calibration is needed
• Tolerant to noisy workplace
• Results:
  o Shows good task generalisation
  o Achieved good success rate and training efficiency: Outperforms the baseline methods with AE-based and VAE-based 

Collab AI developed capabilities for natural dialogue interaction which allows humans to learn from robots by asking questions, and querying the system about tasks, objects and robot status.

• Reduce training costs for new team members, by answering basic questions from them
• Reduce time for switching between hands-on work and machine monitoring, where robots can answer questions about machine states
• Achieve high accuracy with a small dataset of dialogue Q&As