Sling2Sim2Real: One-Shot Elastic System Identification for Non-Destructive Slingshot Policy Learning

Technical Summary: Sling2Sim2Real

Problem Statement
Elastic object manipulation (EOM) presents significant challenges due to high-dimensional, nonlinear deformation dynamics and entropic elasticity. While simulation offers a safe and scalable alternative to costly real-world trials for learning manipulation policies (e.g., slingshot tasks), a critical "sim-to-real" gap persists. This gap arises because elastic properties (such as Young's modulus and damping) are often visually indistinguishable, making it difficult to construct physically consistent simulated counterparts. Existing Real2Sim2Real methods typically require extensive real-world interaction data and struggle with the multimodal, highly nonlinear parameter landscapes of elastic objects, leading to poor policy transfer when applied to unforeseen elastic materials.

Methodology
The authors propose Sling2Sim2Real, a one-shot Real2Sim2Real framework designed to identify elastic parameters from a single non-destructive interaction and enable zero-shot policy transfer. The framework operates in two stages:

1. One-Shot System Identification (Real2Sim):

   • Data Acquisition: The system records a single interaction episode where a robot (Franka Emika Panda) grasps, stretches, and releases an elastic sling without launching a projectile. Data includes synchronized end-effector velocity, wrist-mounted force-torque sensor readings, and visual observations (point clouds) of the sling deformation.
   • Loss Formulation: To calibrate the simulator, the authors define a composite loss function $L = L_{geom} + L_{force} + L_{work}$.
     • $L_{geom}$: Minimizes visual deformation gaps using Chamfer distance between real and simulated point clouds.
     • $L_{force}$: Aligns restoring forces at each time step.
     • $L_{work}$: A novel term measuring the accumulated mechanical work (force-velocity product) to ensure consistency in energy dissipation and stiffness scale, addressing noise and temporal misalignment issues in sample-wise force matching.
   • Optimization Strategy: The authors employ a hierarchical, covariance-informed optimization approach to navigate the complex parameter space (estimating Young's modulus, Poisson's ratio, elasticity damping, dynamic friction, and damping scale):
     • Global Exploration: Differential Evolution (DE) generates a population of candidate solutions.
     • Local Refinement: The top candidates are processed via Non-Maximum Suppression (NMS) to identify distinct regions. Covariance matrices are estimated from these regions to initialize multiple parallel CMA-ES (Covariance Matrix Adaptation Evolution Strategy) solvers. This allows for efficient local optimization that respects parameter correlations.

2. Policy Learning (Sim2Real):

   • The calibrated simulator is used to train a reinforcement learning (RL) policy using Proximal Policy Optimization (PPO).
   • The slingshot task is formulated as a one-step Markov Decision Process (MDP) where the agent selects a draw position to hit a target.
   • The learned policy is transferred directly to the real-world system without further adaptation or fine-tuning.

Key Contributions

• One-Shot Identification: A framework that identifies five critical elastic parameters from a single non-destructive interaction, eliminating the need for extensive real-world data collection.
• Covariance-Informed Optimization: A novel two-stage optimization architecture that leverages parameter covariance estimated from DE candidates to guide multi-start CMA-ES solvers, improving convergence in high-dimensional, nonlinear spaces.
• Work-Based Loss: The introduction of a mechanical work loss term ($L_{work}$) that complements geometric and force losses, significantly improving the calibration of damping and stiffness in elastic materials.
• Zero-Shot Transfer: Successful demonstration of training a policy entirely in simulation and deploying it on a real robot to hit targets with diverse, unforeseen elastic bands.

Experimental Results
The method was evaluated on a slingshot task using three distinct rubber bands (soft, medium, and stiff) and varying target distances.

• Accuracy: Sling2Sim2Real consistently achieved the lowest landing distance errors across all band types and target distances compared to state-of-the-art baselines (including standard Real2Sim2Real, Package SI, and simple DE/Bayesian/Adam optimizers).
• Robustness: Baseline methods frequently failed to launch projectiles or produced errors marked as "N/A" (outside the target area), particularly with stiffer bands. Sling2Sim2Real successfully launched projectiles in all tested configurations.
• Ablation Studies:
  • Removing the covariance-informed multi-start initialization or the work-based loss ($L_{work}$) resulted in significantly higher landing errors, confirming the necessity of both the optimization strategy and the specific loss formulation.
  • The composite loss function yielded lower calibration losses than geometry-only approaches, especially for high-damping materials.

Significance and Claims
The paper claims that Sling2Sim2Real addresses the fundamental challenge of sim-to-real transfer for elastic objects by enabling accurate policy learning with minimal real-world interaction. By treating the elastic object manipulation task as a one-step decision process and utilizing a sophisticated system identification pipeline, the authors demonstrate that robust zero-shot transfer is achievable even when elastic properties are visually indistinguishable and highly nonlinear. The work highlights the importance of incorporating haptic feedback (force and work) alongside visual data and utilizing covariance structures in optimization to overcome the limitations of current Real2Sim2Real paradigms.