Expanded Simulation Improves Robot Quality

Tuong Anh Ens

Go West Robotics expands simulation to improve algorithms and inform design

Overview

Bossa Nova Robotics (BNR) develops retail robots for grocery stores like Walmart and Albertsons. Like all robotics companies, they face pressure to add robot features and to scale their fleet while maintaining high standards of quality.

When developing robots, BNR tests every code change in simulation. The BNR simulator provides a real world model of a retail store and allows BNR to test navigation, mapping and obstacle avoidance. The tests run through dozens of scenarios to confirm that a robot takes the right path, achieves its goals and meets performance metrics.

Go West Robotics expanded the simulator and made a comprehensive assessment of the tests running in simulation. We used our robotics experience to examine scenarios, discovered problematic scenarios and added more scenarios. The expansion of test cases helped the engineering team better understand robot capabilities. We fine-tuned the simulator; setting up metrics correctly, and making sure minimum requirements were met. By making the simulator more like the real world, roboticists could ensure robots performed as expected. Expanding simulation scenarios helped find issues in development, so they were resolved early, before robots were placed in production.

Approach: Real world investigation

Go West studied the BNR robots on location at a retail store. Armed with an understanding of the BNR simulator, we set out to tackle some of the known common problems with mobile robots. For example, robots periodically get stuck and fail to understand their location. Real world investigations allowed us to add scenarios to the simulator like navigating around pillars, small spaces and other specialized obstacles.

In-depth algorithmic assessment

Go West worked with BNR engineers to understand specific algorithms. We drilled down to the details, building tests for positive and negative outcomes. This allowed us to test what is specifically beyond a robot’s capabilities (expected failure). For other tests, we calibrated thresholds for success right up to the point of failure. These test results transform abstract understanding of robot behavior to concrete, useful knowledge. This knowledge informs design and improves algorithms, helping BNR build a robust product.

Results

Improved simulation results in a more robust end product. Difficult problems are thoroughly tested and behavior that was previously unclear becomes transparent and useful. With good simulation, issues are found early and addressed early, saving valuable time and resources before a robot is put into production.