CASE STUDY
How Roborax built a deformable-item picking dataset that improved warehouse robot policy success rate from 61% to 84%.
A warehouse automation company had a robot picking policy that performed well on rigid items but degraded significantly on deformable goods — polybags, pouches, and soft-packaged items that made up 40% of their target SKU range. The policy had been trained almost entirely on rigid-item demonstrations and had never seen the deformation dynamics that real deformable items exhibit.
Deformable-item manipulation is one of the hardest problems in warehouse robotics. The data challenge is that deformable items behave differently every pick — no two polybag picks are identical. The model needs to learn from a wide distribution of deformation states, grasp angles, and item weights. Generating this distribution requires operators who understand the task well enough to intentionally vary their approach.
Roborax ran a structured deformable-item data collection program using a dedicated operator team trained on the client’s robot. Over six weeks, we collected 8,000 picking demonstrations across 47 SKU types, with deliberate variation in approach angle, grip point, item fill level, and shelf position. Tactile sensor data was synchronized with vision and joint state throughout.
Policy success rate on deformable items improved from 61% to 84% after retraining on the Roborax dataset — a 23 percentage point improvement that moved the product from blocked to production-deployable.
Related: Warehouse and logistics solutions — Case studies.
The most important lesson from this program was that deformable-item data collection requires intentional variation by design — it cannot be achieved through passive capture. Operators need to be trained to deliberately vary their approach, and the data collection protocol needs to specify the dimensions of variation explicitly. Without this, a large dataset of deformable-item picks will be large but not diverse, and a model trained on it will be brittle. Related: Warehouse and logistics solutions — Case studies.
