Denmark tests humanoid robots in supermarkets

Denmark is testing humanoid robots in its supermarkets to reduce the workload of night staff. The initiative involves a startup called Hive Robots, which has been testing Unitree humanoid robots in a Føtex store near Copenhagen since June.

Robots guided remotely for restocking

The robots are slower than employees and are still guided remotely. Their goal is to restock certain aisles independently by early 2027. According to a study published in July, their best success rate on a reconstructed aisle currently peaks at 42 percent.

Training the robots with virtual reality

At Hive Robots, Roger Sala uses a virtual reality headset and controllers to guide a humanoid through the aisles of a closed supermarket. Each recorded movement is used to train the machine. The store is in Lyngby, north of Copenhagen, and belongs to Føtex, a chain of the Salling group, the largest Danish retailer.

Challenges in recruiting for restocking

Føtex says restocking absorbs 40 to 50 percent of staff working time, and the chain has difficulty recruiting young people for these tasks. Simon Kudsk, its operations director, envisions "one or two humanoids in each store" and estimates it will take up to five years to equip the network.

Collaboration with Technical University of Denmark

In July, the team published a study with the Technical University of Denmark (DTU) detailing its method, called DEED. The tests take place on a replica aisle with shelves, a trolley, and a box of chip packets. The robot, a Unitree G1-Edu, is piloted by GR00T N1.6, the Nvidia model for humanoids.

Improving success rate through data refinement

The researchers revised the data to improve the robot’s performance. They kept the moments when the operator corrects an error, allowing the robot to learn to correct itself. Visual masks direct the model’s attention to the empty shelf location, and a filter smooths the movements. The success rate then rises to 32 percent.

Learning from its own experience

In a second phase, the robot learns from its own experience. Its successes, failures, and human corrections are reintegrated into the training. After one iteration, the success rate rises to 42 percent, but the robot stops after a single packet instead of returning to its starting position.

Real-time failure detection

A tool detects in real time when the robot deviates from what it has learned, allowing the anticipation of failures. The authors conclude that bridging the gap between the laboratory and the store depends more on system integration and data care than on a new AI architecture.

Long-term goals for robot deployment

Jakob Sand, founder of Hive Robots, compares current robotics to what conversational AI was two or three years ago. His startup aims for autonomous night restocking of certain products by early 2027, before tackling more complex tasks like detecting missing or misshelved products.