Turning 'Bad' Robot Data Into Precision Gains for VLA Models

Training robots to perform high-precision manipulation—tasks requiring sub-millimeter accuracy—usually demands large amounts of carefully collected, task-specific demonstration data, often gathered through slow and expensive teleoperation. A new paper posted to arXiv, as reported by arXiv (cs.RO), proposes a way to cut that cost by making use of data that would normally be thrown away.

The method, called ε4P (short for "Imperfection for Precision"), targets vision-language-action (VLA) models, which map visual input and language instructions directly to robot actions. Instead of relying solely on pristine, high-quality demonstrations for the exact task at hand, ε4P "upcycles" two imperfect data sources: low-precision recordings of the target task itself, and high-precision recordings of different, mismatched tasks.

The key insight is not simply mixing these two imperfect datasets together during training. Rather, ε4P is built around flow-matching, a generative modeling technique that produces actions by gradually denoising a trajectory. The authors control which data source contributes at which stage of that denoising process: low-precision, target-task data is used when noise levels are high, helping the model retain high-level context about what the task actually is, while high-precision data from mismatched tasks is used when noise levels are low, transferring fine-grained action precision even though it comes from a different task altogether.

The team tested the approach with real-robot experiments spanning both sub-millimeter, high-precision tasks and coarser-grained tasks. Two results stood out. First, adding these otherwise-discarded imperfect data sources improved policy performance by as much as 31.7 percentage points compared to not using them. Second, and perhaps more notably for practical deployment, ε4P could substitute an equivalent volume of expensive, task-specific, high-quality data with only an average 4.2 percentage point drop in performance—suggesting that a large share of costly precision-focused data collection could potentially be avoided.

The authors frame ε4P as a step toward a more scalable paradigm for precision manipulation, where heterogeneous and imperfect data—rather than being discarded as unusable—is systematically routed to different parts of a training pipeline based on what it's good for: task context versus action precision. If this approach generalizes across more tasks and robot platforms, it could meaningfully lower the data-collection barrier that currently limits how quickly high-precision manipulation policies can be developed and deployed.

The paper, submitted by author Wei Hao, is a 9-page report with 5 figures, and the authors note that additional details are available on a project page linked from the arXiv listing. As with many recent VLA papers, the practical impact will depend on how well the method scales to more complex tasks, longer horizons, and different robot embodiments beyond the settings tested here.