ActivePusher

Active Learning and Planning with Residual Physics for Nonprehensile Manipulation

Zhuoyun Zhong Seyedali Golestaneh Constantinos Chamzas

Worcester Polytechnic Institute · ELPIS Lab

Accepted to ICRA 2026

ICRA 2026 Best Student Paper Award ICRA 2026 Best Paper in Planning and Control Finalist HRF 2026 Best Student Paper Award

Active learning

Collecting informative interactions improves pushing dynamics model efficiently.

Active planning

Uncertainty-aware planning favors reliable actions during task execution.

Learn where uncertain. Plan where reliable.

A UR10 robot pushes household objects to target regions and table edges using ActivePusher
ActivePusher learns object dynamics from targeted interactions, then uses model uncertainty to construct more reliable manipulation plans.

Planning with learned dynamics can unlock versatile nonprehensile manipulation, but collecting robot data is expensive and prediction errors can compound over long horizons. ActivePusher joins residual physics, uncertainty-aware active learning, and kinodynamic planning in one framework: the robot practices the pushes that teach it the most, then favors high-confidence actions when it is time to act.

The result: more data-efficient model learning and higher planning success in both simulation and real-world pushing tasks—without high-fidelity simulation, large offline datasets, or human demonstrations.

One uncertainty signal, two roles

ActivePusher framework showing active learning and active planning driven by estimated model uncertainty
Estimated model uncertainty guides the system toward informative actions during learning and reliable actions during planning.

Active learning

ActivePusher estimates epistemic uncertainty with the neural tangent kernel and uses BAIT to select a batch of skill parameters with high expected information gain. Each real interaction is chosen to improve the dynamics model efficiently.

Active planning

The same uncertainty estimate biases the kinodynamic planner toward controls in well-explored parts of the skill space. Plans are built from actions the learned model can predict with greater confidence.

Push to region

Push-to-region trial 1

Push-to-region trial 2

Push-to-region trial 3

Push-to-region trial 4

The robot plans and executes multi-step pushes that move a mustard bottle into a target region while accounting for model uncertainty.

Push to edge for grasping

Push-to-edge trial 1

Push-to-edge trial 2

Push-to-edge trial 3

Push-to-edge trial 4

With obstacles on the table, the planner moves a cracker box to a reachable edge so the robot can complete the task with a grasp.

Residual physics for low-data learning

A coarse analytical pushing model supplies a useful physical prior. A neural network learns only the residual correction between that approximation and observed object motion, retaining physical structure while adapting to object- and contact-specific behavior.

Plots comparing active and random learning with residual-physics and fully learned models
Combining residual physics with active data selection reaches strong predictive performance with substantially fewer interactions than the baselines.

Uncertainty-aware planning improves execution

Active sampling steers the planner toward actions with lower model uncertainty. Across planning conditions, this produces more executable plans, improves success, and reduces the gap between planned and observed object motion.

Planning success rate and execution error plots for ActivePusher and baseline methods
Active planning improves success rate and execution accuracy for the learned pushing models.

Acknowledgments

This work was supported in part by NSF CRII Grant No. 2451108, an Amazon WPI Robotics Engineering gift, and Worcester Polytechnic Institute funds.