ThoughtForge and Active Inference will Revolutionize Robotics

In our first post, we announced the launch of the ThoughtForge private beta and our vision to free humans from dull, dirty, and dangerous jobs by bringing common sense to robots. We are achieving this vision by making it absurdly easy to build and manage robust and truly autonomous robotic control for the real world.

A Robotic control system controls the movement and functions of a robot. A robot can have one or more control systems for various purposes. Similar to the human brain, control systems process data from the robots’ senses (or sensors) to understand the environment, calculate the appropriate action, and tell the actuators to perform the calculated action. The operating environment and dynamics of the robot (motor torque, data from sensors, etc.) vary with the time, therefore, it is important for robotic control systems to always be accurate, adaptive, robust, responsive, and safe in their responses.

Simple Robotic Control System

Currently, robotic control is primarily developed using Inverse Kinematics solvers (IK solvers), Deep Reinforcement Learning (DRL), or a combination of both. IK solvers require the robots movement and robots' operating environment to be precise to achieve repeatable results. DRL uses simulations to train the robot for a specific task and assumes the simulated world and real world are the same. Therefore, existing robotic control and robots lack the robustness, accuracy, and adaptability needed to perform safely in dynamic real world environments. They always fail when faced with uncertain or unseen situations. A great example that highlights this is Amazon which has faced significant challenges while trying to get their drones to operate in unobstructed or clear skies (Source: Verge, Bloomberg). Even companies like FedEx and Walmart have not been able to get robots to satisfactorily perform tasks in their sorting facilities, warehouses, and stores.

This is why robots primarily continue to be limited to known and constant environments such as single item manufacturing lines. Any modification to these environments requires retraining robots which involves costly simulations changes, unceasingly growing data, and exponentially high training costs and time.

ThoughtForge changes all this using the power of Active Inference. It enables customers to build optimized, robust, and accurate robotic control models using their existing simulation and data sources.

You might wonder what Active Inference is, what benefits ThoughtForge’s implementation of Active Inference provides, and how ThoughtForge will revolutionize Robotics? Let's dig in!

Active Inference

Developed by Prof. Karl Friston who is an advisor to ThoughtForge, Active Inference is the theory of how all biological systems (or agents) learn. It explains how an agent learns from, interacts with, and perceives the environment it is in. It is based on the Free-Energy Principle, a.k.a, FEP (or Surprise Minimization).

The Free-Energy Principle is based on the premise that an agent, in order to maintain its existence and avoid death, must ensure that important parameters, like body temperature or blood oxygenation, don’t deviate too much from the norm, i.e. are not surprising.

A book has recently been published on Active Inference, and is available via Open Access:

https://mitpress.mit.edu/books/active-inference

Learning & Surprise

Agents learn by interacting with their environment and build upon their learning over time to build a generative model of their environment to predict future states. An agent can minimize Surprise (mis-prediction) in two ways: either by an agent updating its model about the world (perception) or via an agent acting on the world to move them into a less-surprising state (action).

Active Inference

However, agents can never predict the future with perfect accuracy, and uncertainty will always be present. Therefore, the learning goal is to minimize uncertainties for as long as possible, within the constraints of the environment. The agents always aim to minimize the upper bound of the surprise (or free energy) by using both their knowledge (or beliefs) from previous learnings and sensory inputs. For robots that means reducing error and making better decisions to achieve the required results, while respecting the boundaries of the environment they are operating in.

Active Inference provides a principled account of epistemic exploration and intrinsic motivation--incorporating uncertainty as a natural part of belief updating. This provides the agent with a benefit of natural trade off between epistemic exploration and pragmatic behavior. (Schwartenbeck P, et al. (2013), “Exploration, novelty, surprise, and free energy minimization”).

To learn more about how Active Inference will unlock new capabilities for Robots, check out this recent paper “How Active Inference Could Help Revolutionise Robotics" https://www.mdpi.com/1099-4300/24/3/361

Benefits of ThoughtForge’s Active Inference powered platform

ThoughtForge was born after 15 years of research and experimentation, and is the sole product on the market that creates accurate and robust autonomous robotic control(s) for the real world. Today, Data Scientists and Robotic Control Engineers can use ThoughtForge to create robotic control models for tasks like object manipulation, material handling, inspection, packaging, palletizing, and many more in dynamic environments of factories, manufacturing facilities, and warehouses. Our fully managed service provides the following advantages:

1. Low Cost / High Sample Efficiency: We reduce training costs by over 99.999% and are 90,000 times more sample efficient when compared to Deep Reinforcement Learning (DRL). The inherent advantage of Active Inferences’ structured generative model is that it enables highly sample-efficient (cost-effective) learning because it builds itself through environmental feedback and has a structured understanding of the complicated rules of cause and effect.

2. Adaptive and Robust: Our models are robust in completing the tasks, adapt without any intervention and adjust on-the-fly to account for any differences encountered in the real world. On the other hand, DRL models and systems fail when they encounter scenarios not seen in training/simulations.

Models built using ThoughtForge’s patent-pending framework use a real-time inference process, adapt to new situations and scenarios and can be built and evaluated in seconds. ThoughtForge is enabling a new paradigm for rapid design and iteration of robotic control models.

Why ThoughtForge and Active Inference will revolutionize Robotics?

ThoughtForge makes it convenient and fast to create robotic control models for low level tasks. Users can easily connect their existing data sources using our API or SDK to train their model or pick from our library of battle tested models. With ThoughtForge, robotic control models for simple tasks like palletizing, pick and place, object manipulation that take 7+ months, can be done in a matter of minutes. Our robotic control models are also fully compatible with the Robot Operating System (ROS), your existing AI/ML models, and robots.

ThoughtForge Platform

ThoughtForge models adapt to changes in the environment, learn on the fly, and apply these learnings to achieve results without sacrificing performance or safety. All of this – in real time! ThoughtForge models also share their learnings with other robots in a fleet. We bring real time optimization to your entire fleet of robots. A ThoughtForge model developed for a specific robot and a specific task works out of the box when deployed on a similar robot for a similar task.

ThoughtForge is currently in a private beta, and would love to help your Enterprise Development and/or Data Science team build, deploy, and manage autonomous robotic controls that works in the real world.

Get in touch and let us help you with your use cases at info@thoughtforge.ai.