VERSES unveils robotics architecture that works without model pre-training

Cognitive computing company, VERSES AI Inc., which specializes in next-generation agentic software systems, unveiled more details on the development of its robotics model.

The Vancouver-based company said that the VERSES robotics architecture accomplished typical household tasks of tidying the room, preparing the groceries and setting the table, better than other robotics models. And, it accomplished the tasks without any pre-training.

Robotics model adapts to new situations

VERSES said that robots often perform well on scripted tasks but can freeze when faced with new situations; even something as simple as a misplaced box can halt progress.

Newer approaches can be more flexible but require huge amounts of training to be effective, which the company said makes existing robotics systems difficult to use in real-world applications where new situations constantly arise.

VERSES said that challenges like this are well-suited to its models’ abilities for quickly adapting to their environment.

“Currently robotics systems are often brittle, and need huge amounts of training data, which makes them expensive and prone to going wrong.” said Sean Wallingford, former CEO and President of Swisslog. “For instance, if you bring a robot to a new factory or ask it to do a different job, it will need a lot of re-training and may not be reliable. VERSES breakthroughs are exciting, because they offer an alternative approach. If we can deploy robots without training, they will be viable in a wide range of activities, from factories and warehouses to domestic and commercial applications.”

In a published paper entitled, Mobile Manipulation with Active Inference for Long-Horizon Rearrangement Tasks, written by members VERSES’ research lab, the robotics model is compared to a deep learning alternative in three tasks:

• Tidying a room
• Preparing groceries
• Setting a table

The VERSES robotics model achieved a success rate of 66.5% across these tasks while the previous best alternative had a success rate of 54.7%. The VERSES robotics model also requires no training, whereas the report found that the other robotics model required 1.3 billion steps to pre-train several skills across the three tasks.

“I believe that by combining our world-modeling and our active inference capabilities, we’ve shown robots can think on their ‘feet’ - navigating and completing complex tasks without months of costly training,” saod Hari Thiruvengada, CTO of VERSES. “Our breakthrough has the potential to transform how robots operate across industries, from factories and warehouses to homes and public spaces, potentially unlocking a new era of truly adaptive, reliable automation.”

VERSES said it plans to present the paper at the International Workshop on Active Inference later in 2025.

Robot model utilization

VERSES said robots generally fall into two categories: drive-by-wire or deep learning.

Drive-by-wire means everything is pre-programmed, whereas deep learning relies on vast amounts of data for training. For example:

• Autonomous Guided Vehicles (AGVs)
  • The drive-by-wire approach breaks down if anything is out of place.
    • For instance, a human might program a robot to move an object to a location by providing a very detailed list of tasks in the form of a plan (e.g. “pick an item from the shelf and place it on the shelf”) down to the specific movements needed by each joint on a robot’s arm.
    • However, factories and homes are always changing. Robots often struggle to adapt, which can cause them to stop or work very slowly.
    • To overcome their inherent limitations, robotics environments are often controlled. For instance, robots may be placed in a cage or in factory areas where no humans are allowed. This practice greatly reduces the robots' usefulness.
• Deep learning approaches
  • Deep learning approaches, by contrast, are trained on vast volumes of data, so that they are more flexible.
    • However, these methods struggle with situations outside their training. Simple issues, like a bottle falling over or a chair being out of place, can confuse and paralyze the robot as it cannot adapt.
    • For instance, a robot replenishing a production line in a factory may not be adaptable enough to switch aisles when its initial route is blocked. Or if it is placing a bowl on a table, it may not be able to adapt to existing objects such as a wine glass, even if it has placed them there itself.

VERSES offering tackles adaptability

The company said it has solved this problem of adaptability.

When a human needs to get a drink in a new apartment, they don't execute by having practiced this task in hundreds of different apartments; they can adapt because they have a model of how the world works. This allows humans to figure out that they need to open the refrigerator and grab a bottle.

VERSES said its technology equips robots with a world model, allowing them to execute three tasks in different apartment layouts.

VERSES models, similar to its work on the AXIOM digital brain, don’t require any pre-training and instead just adapt by exploring the environment.

VERSES said its models consist of three modules working together:

• Vision: Taking pixels and turning them into understanding, as well as mapping the room it is in.
• Planning: It can take a task such as setting the table for dinner and plan out all the subtasks (e.g., opening a drawer, and putting cutlery on the table) without needing detailed instructions
• Control: Translating these into all the specific movements of the robot and its arm.

At each stage, VERSES said its system can adapt. For example, it can cope with unexpected objects in its way, or needing to pick up something it has dropped.

VERSES also said that most importantly, its model needs no training. All the VERSES model news is basic knowledge such as its arm resting pose when idle or how much resistance the arm will gets from obstacles. By contrast, VERSES said the baseline model of other systems requires extensive offline training of 6,400 episodes per task and 100 million steps per skill across a total of seven skills, such as picking up an object or opening a fridge.

The company said that use cases for this work could include moving inventory around factories and warehouses.