ActivePusher: Active Learning and Planning with Residual Physics for Nonprehensile Manipulation

TL;DR

This work proposes ActivePusher, a novel framework that combines residual-physics modeling with uncertainty-based active learning, to focus data acquisition on the most informative skill parameters, and seamlessly integrates with model-based kinodynamic planners.

Abstract

Planning with learned dynamics models offers a promising approach toward versatile real-world manipulation, particularly in nonprehensile settings such as pushing or rolling, where accurate analytical models are difficult to obtain. However, collecting training data for learning-based methods can be costly and inefficient, as it often relies on randomly sampled interactions that are not necessarily the most informative. Furthermore, learned models tend to exhibit high uncertainty in underexplored regions of the skill space, undermining the reliability of long-horizon planning. To address these challenges, we propose ActivePusher, a novel framework that combines residual-physics modeling with uncertainty-based active learning, to focus data acquisition on the most informative skill parameters. Additionally, ActivePusher seamlessly integrates with model-based kinodynamic planners, leveraging uncertainty estimates to bias control sampling toward more reliable actions. We evaluate our approach in both simulation and real-world environments, and demonstrate that it consistently improves data efficiency and achieves higher planning success rates in comparison to baseline methods. The source code is available at https://github.com/elpis-lab/ActivePusher.

Figures (7)

• Figure 2: Neural network (NN) with residual physics architecture. The network takes both the control parameters and the output of the physics model to predict residuals, which are then added to the physics-based output to produce the final prediction.
• Figure 3: a) ActivePusher quantifies the model uncertainty of the learned model (\ref{['subsec:uncertainty']}). b) During the learning phase, activePusher chooses the most informative push to apply to increase the learning efficiency (\ref{['subsec:active_learning']}). c) During planning the most reliable pushes are chosen to maximize the task success rate (\ref{['subsec:active_planning']}).
• Figure 4: Experiment Setup
• Figure 5: Skill Learning results show SE2 prediction error for 2 real objects, and 4 simulated objects from ycb: Banana, Mug, Cracker Box and Mustard Bottle. The active learning methods outperform random data collection, and models with residual physics perform better in low-data regime.
• Figure 6: 2D schematic of the planning tasks