Technical Summary: Periodic Robust Robotic Rock Chop via Virtual Model Control

Problem Statement

Robotic cutting is a complex, contact-rich manipulation task characterized by unpredictable material fracture, deformation, and significant geometric uncertainties (e.g., knife shape, cutting board height, and food properties). Existing approaches face two primary limitations:

1. Learning-based methods require extensive interaction data to converge, posing safety risks and generating food waste during training. They often struggle to generalize to object properties or tool geometries outside the training distribution.
2. Model-based and trajectory-tracking methods rely on precise kinematic models of the knife and environment. They typically fail when faced with unmodeled geometric constraints or when the robot encounters contact conditions different from the pre-planned trajectory. Furthermore, generating rhythmic "rock-chop" motions via predefined paths demands exact characterization of environmental constraints, which is often unavailable in practical settings.

The core challenge is to achieve robust, rhythmic cutting without relying on high-fidelity modeling, precise environmental sensing, or pre-computed trajectories.

Methodology

The authors propose a Virtual Model Control (VMC) framework that generates robust, self-sustained rock-chop motions through physically structured interaction rather than explicit trajectory planning.

Virtual Mechanism Design

The controller employs a switched virtual mechanism that shapes the robot's behavior through dynamic interaction with the environment:

• Virtual Mass and Springs: A virtual mass slides along the rocking direction (y-axis), connected via a spring-damper to the knife tip ($p_1$). This ensures continuous contact with the cutting board while allowing free sliding.
• Switched Topology: The system alternates between two virtual mechanical topologies corresponding to the cutting and raising phases.
  • A second spring-damper connects the knife handle ($p_2$) to a control point ($p_b$).
  • $p_b$ switches between two reference points ($r_{2,1}$ and $r_{2,2}$). Switching to $r_{2,1}$ drives the knife down (cutting), while switching to $r_{2,2}$ lifts the knife (raising).
  • Transitions are triggered by the knife's orientation (spine alignment with the board) and position relative to reference points, modulated by a smoothing mass-spring-damper system to ensure continuous dynamics.
• Plane Constraints: Four additional virtual points on the knife are connected to virtual masses constrained to a desired slicing plane ($r_S$) via springs and dampers. This enforces planar motion to ensure clean, even cuts.
• Nonlinear Elements: All virtual springs utilize a saturating nonlinear characteristic ($f = \sigma \tanh(k|z|/\sigma) \frac{z}{|z|}$) to limit excessive forces and enhance safety.

Theoretical Analysis

The authors provide an energy-based analysis demonstrating that the system settles into a stable limit cycle:

• Energy Balance: The periodic motion emerges from an equilibrium between energy injected via reference switching and energy dissipated through virtual damping, robot internal dissipation, and environmental resistance (food/board interaction).
• Stability Certification: Stability is analyzed via the contractivity of the associated Poincaré map. Numerical linearization of the return map shows that eigenvalues remain within the unit circle for a wide range of parameter variations, confirming asymptotic stability of the limit cycle.

Key Contributions

1. VMC-Based Limit Cycle Controller: A novel controller design that generates robust, self-sustained rock-chop motions through controlled mechanical switching and regulated energy injection, eliminating the need for explicit trajectory planning.
2. Theoretical Stability Proof: Analytical and experimental verification of the stability and periodicity of the cutting motion using passivity theory and Poincaré map analysis.
3. Comprehensive Experimental Characterization: Evaluation of cutting performance across five vegetable types (spring onion, cucumber, courgette, carrot, potato) with varying thicknesses (1–6 mm) and speeds.
4. Robustness Validation: Demonstration of the controller's ability to adapt to perturbations in knife geometry (two different knives) and cutting board height (up to 40 mm variation) without re-tuning.
5. Platform Independence: Successful implementation and retargeting on two distinct robotic platforms: the Franka FR3 and the RT Corp Sciurus 17 humanoid robot.
6. Online Adaptation Study: A preliminary investigation into online parameter adaptation (adjusting stiffness $k_2$ based on food separation success) to handle changing material properties.

Experimental Results

Experiments were conducted using a Franka Research 3 arm equipped with a force-torque sensor and a rigidly attached knife.

• Performance: The system achieved sub-millimeter slice accuracy for thicknesses ranging from 1 mm to 6 mm at a rate of nearly one cut per second.
• Robustness to Geometry: The controller maintained stable rocking motion and successful cuts with two different knives (different blade curvatures) using a single set of parameters. In contrast, baseline impedance and hybrid force/position controllers relying on pre-defined trajectories failed to maintain contact or produced incomplete cuts when the knife geometry changed.
• Robustness to Environment: The controller successfully adapted to cutting board height variations from 0 to 40 mm. Baseline controllers failed at higher board heights due to excessive contact forces triggering protective stops or loss of contact.
• Parameter Sensitivity: Sensitivity analysis (via MuJoCo simulation) showed that the system remains stable across ±20% variations in most parameters. The only failure case occurred when switching thresholds were too close, causing rapid oscillation between states.
• Comparison: Compared to standard impedance control and hybrid force/position baselines, the proposed VMC approach demonstrated superior robustness to geometric mismatches and environmental uncertainties, as it does not rely on tracking a nominal trajectory.

Significance and Claims

The paper claims that Virtual Model Control provides a robust design framework for complex, contact-rich manipulation that significantly reduces reliance on high-fidelity modeling and computational overhead.

• Structural Robustness: The authors argue that robustness is derived from the structure of the controller itself. The behavior is shaped by physically meaningful interaction dynamics rather than a precomputed trajectory or a learned mapping.
• Generalization: Unlike learning-based approaches that require extensive data or model-based approaches requiring precise geometry, this method requires only approximate task-level initialization (e.g., approximate object location and knife orientation) and minimal parameter tuning.
• Limitations and Scope: The authors modestly state that the work focuses on the control layer for generating robust periodic cutting behavior under uncertainty. It does not solve the entire autonomous food-preparation pipeline (e.g., pick-and-place, food holding, or real-time vision integration), though it is designed to be integrated into such a pipeline. The online adaptation study is presented as a preliminary step toward more advanced adaptive capabilities.

The work demonstrates that a physically structured controller can effectively handle the uncertainties inherent in robotic cutting, offering a viable alternative to trajectory-tracking and data-intensive learning methods for rhythmic, contact-rich tasks.