Deformable objects pose planning challenges because their state space is high-dimensional, contacts are intermittent, and material properties vary. These causes make pure open-loop planners brittle and motivate strategies that trade exactness for speed and robustness. Ken Goldberg University of California Berkeley has long emphasized practical systems for cloth and rope where perception-driven feedback reduces uncertainty. Sergey Levine University of California Berkeley and Aude Billard École Polytechnique Fédérale de Lausanne advocate learning methods that capture complex dynamics without full analytic models, enabling real-time adaptation.
Core technical strategies
Efficient real-time planning typically combines model-based and learning-based elements. A practical approach uses reduced-order models to represent cloth or soft bodies with far fewer degrees of freedom, feeding a fast model-predictive control loop that replans on sensory updates. Approximate models are acceptable when paired with robust feedback, because the planner need only predict a short horizon. Complementary strategies use learning-based policies trained in simulation and refined on real hardware through online adaptation, a workflow shown to accelerate deployment by groups working on sim-to-real transfer.
Perception and sensing matter: fusing depth vision with tactile sensing allows planners to correct for unseen deformation in real time. Task decomposition into primitives such as grasp, stretch, and align reduces planning complexity and lets controllers reuse precomputed motions. When dynamics are highly uncertain, fast parameter identification or Bayesian online estimation can update the planner’s internal model while executing, maintaining responsiveness without exhaustive search.
Relevance and consequences
Efficient real-time methods enable automation in apparel manufacturing, surgical assistance, and household robotics, with social consequences for labor and skill demands. In regions with strong textile industries, automation can raise productivity but may displace manual jobs unless retraining programs accompany adoption. Environmentally, better handling of deformable materials can reduce waste in sorting and recycling operations by enabling accurate manipulation of flexible items. Cultural practices that govern how fabrics are handled vary by territory, so systems must adapt to local conventions and material varieties.
Adopting hybrid, perception-rich planners that prioritize short-horizon accuracy, reuse of primitives, and online learning produces practical, real-time performance for deformable object tasks while acknowledging trade-offs between generality and speed.