Online Material Estimation for Conditioned Diffusion Policy in Shaping Deformable Linear Objects

Technical Summary: Online Material Estimation for Conditioned Diffusion Policy in Shaping Deformable Linear Objects

Problem Statement

The manipulation of Deformable Linear Objects (DLOs), such as cables, tubes, and strings, presents significant challenges for imitation learning. Unlike rigid objects, DLOs exhibit deformation behaviors that vary drastically based on material properties like stiffness, elasticity, and surface friction. Consequently, a single manipulation policy must generate distinct action sequences for different materials to achieve the same goal shape.

Existing approaches face limitations:

• Model-based methods require explicit physical modeling and re-identification for every new object, struggling with contact-rich interactions.
• Standard imitation learning often assumes a fixed object or relies on simulators to identify object properties, which are unavailable in practical, real-world scenarios where the object identity is unknown a priori.
• Prior conditional policies (e.g., GenDOM) often require dedicated probing actions, reference demonstrations, or simulator-based identification, making them less adaptable to dynamic, real-world settings.

The core problem addressed is how to enable a robot to manipulate unknown DLOs effectively without prior knowledge of their material properties, using only observations gathered during the manipulation task itself.

Methodology

The authors propose a framework comprising two primary components that operate in tandem during execution: a Material Estimation Network and a Material-Task Conditioned Diffusion Policy.

1. Online Material Estimation Network:

   • Input: A time series of the last 10 observations, including RGB images from environment and hand cameras, and robot joint states.
   • Architecture: Visual features are extracted via ResNet18, and joint states via a Multilayer Perceptron (MLP). These are concatenated and processed by a single-layer Long Short-Term Memory (LSTM) network to capture temporal interaction history. An attention pooling module aggregates these features, which are then mapped to a probability distribution over material labels via an MLP.
   • Function: The network predicts the material label of the grasped object online. Crucially, this estimate is re-evaluated throughout the execution of the task, rather than just once at the beginning.

2. Material-Task Conditioned Diffusion Policy:

   • Architecture: A diffusion policy based on a 1D U-Net that generates action sequences through iterative denoising.
   • Conditioning: The policy is conditioned on a global feature vector derived from the current observation (visual and joint states) and one-hot encoded labels for both the task (target groove shape) and the material.
   • Training vs. Inference: During training, the policy uses ground-truth material labels. During inference, the ground-truth label is replaced by the online estimate provided by the material estimation network at every inference step.

The system requires no simulator, no reference demonstrations, and no dedicated probing actions; the material label is inferred solely from the task's own interaction history.

Key Contributions

The paper outlines three main contributions:

1. A Simulator-Free Framework: The proposed method infers material labels directly from task observations and continuously re-estimates them during execution, eliminating the need for pre-specified object identities or simulator-based identification.
2. Performance Validation: Using 480 real-robot demonstrations across four materials (Silicone, Thick acrylic, Thin acrylic, Cotton) and three groove-placement tasks, the authors demonstrate that conditioning on online estimated labels achieves manipulation performance comparable to conditioning on ground-truth labels.
3. Behavioral Analysis: Through detailed analysis of recorded rollouts, the paper shows that the estimator successfully extracts material-related information from manipulation observations and that the diffusion policy responds dynamically to changes in the conditioning signal.

Experimental Results

The framework was evaluated on a UR5e manipulator with a setup involving four DLO materials and three groove-placement tasks (Crank, Vertical S, Horizontal S). Four policies were compared:

• Specialist Policy: Trained separately for each material-task pair (40 demos each).
• Task-Only Policy: Trained on all data but conditioned only on the task label.
• Oracle Policy: Trained on all data and conditioned on ground-truth material labels.
• Estimated Policy: The proposed method, using online estimates.

Key Findings:

• Success Rates: The Estimated Policy achieved an average success rate of 60.8%, closely matching the Oracle Policy (60.0%) and significantly outperforming the Task-Only Policy (45.8%) and the Specialist Policy (41.7%).
• Material Estimation: The estimator achieved high accuracy (up to 96.26% with optimized architecture) when trained on data generated by specialist policies, indicating that specialized trajectories contain clearer material-dependent deformation characteristics.
• Policy Responsiveness: Analysis showed that the policy's commanded velocity changes significantly (up to 2.17x) when the estimated material label changes, confirming the policy actively utilizes the conditioning signal.
• Failure Analysis: The primary failure case occurred with Silicone on the S-Vertical task. Post-hoc analysis revealed this was due to persistent confusion between Silicone and Thin acrylic by the estimator, suggesting that when the estimator cannot distinguish between two similar materials, performance degrades.

Significance and Claims

The paper claims that its framework effectively bridges the gap between the need for material-aware manipulation and the reality of unknown object identities in practical robotics. By demonstrating that a policy conditioned on estimated labels performs as well as one conditioned on ground-truth labels, the authors validate that online material estimation is a viable strategy for DLO manipulation without prior knowledge.

The authors modestly note that while the estimator does not need to be perfectly accurate for the conditioning to be beneficial, persistent confusion between similar materials remains a limitation. Future work is identified as generalizing to unseen materials and potentially replacing discrete labels with higher-dimensional, interpretable representations (e.g., friction, elasticity) to enable explainable action generation.