Teaching Robots to Coexist with Humans
At Config, we believe the hardest part of deploying robots isn’t executing the task — it’s operating in environments where people are present.
Most robot learning today happens in closed environments — human-free workspaces where robots learn to pick, place, and assemble in isolation. This works for isolated automation, where robots are separated from people. However, we don’t believe this is how every real-world environment will evolve. Many tasks are inherently shared, and in practice, they’re often more efficient (in terms of both time and cost) when humans and robots work together.
Yet this simple reality — that people are part of the environment — is largely missing from how robots are trained today, despite being critical for real-world deployment.
We’re working to close this gap: teaching robots to operate reliably with humans around.
What we're building
We've been collecting demonstrations in human-present environments, where a person enters the scene, gives direction, or works together with the robot toward a shared goal. Importantly, these demonstrations are designed to prioritize human-friendly behavior — going beyond task completion to include collision avoidance, appropriate waiting, and yielding in shared spaces.
When we train our models on this data, something interesting happens: human-friendly behaviors emerge on their own. In our recycling sorting task (where the robot sorts cans and the human sorts other materials), for instance, the robot learns to pause when a person reaches into the bin and wait while someone sorts alongside it — none of which we explicitly programmed. This isn't driven by explicit safety rules, but by patterns in the demonstrations — where the robot repeatedly observes when to pause, yield, or stay out of the way of a person. These behaviors surface naturally because the demonstrations showed not just how to complete the task, but how to do it alongside a person, with human actions consistently taking precedence.
Models trained the conventional way — on human-absent data — simply don't develop these behaviors. They collide with a person in the way, and they don't wait when someone is nearby.
What’s Next?
What's Next? We see this as an important step toward robot foundation models that are ready for the places where people actually live and work — homes, factories, care facilities. If you're interested in our data, models, or just want to talk, we'd love to hear from you at https://forms.config.inc/contact.
We're also planning to open-source the dataset in the near future, so stay tuned.
