Quality Over Quantity: Curating Contact-Based Robot Datasets Improves Learning

TL;DR

This work challenges the notion that more data universally aids robot learning, focusing on contact-based tasks where information content varies across samples. It introduces a contact-aware Fisher information measure to rank and curate data, and extends optimal experimental design to generate information-rich contact demonstrations for the ContactNets framework. Empirical results show that Fisher-ranked data can improve learning efficiency, reduce variance, and sometimes enable deterministic learning, especially under limited data. The proposed data-curation and generation methods offer a practical path to more data-efficient, contact-informed robot learning with potential applicability beyond ContactNets.

Abstract

In this paper, we investigate the utility of datasets and whether more data or the 'right' data is advantageous for robot learning. In particular, we are interested on quantifying the utility of contact-based data as contact holds significant information for robot learning. Our approach derives a contact-aware objective function for learning object dynamics and shape from pose and contact data. We show that the contact-aware Fisher-information metric can be used to rank and curate contact-data based on how informative data is for learning. In addition, we find that selecting a reduced dataset based on this ranking improves the learning task while also making learning a deterministic process. Interestingly, our results show that more data is not necessarily advantageous, and rather, less but informative data can accelerate learning, especially depending on the contact interactions. Last, we show how our metric can be used to provide initial guidance on data curation for contact-based robot learning.

Figures (5)

• Figure 1: Informative contact-data improves learning. Here, we show demonstrations of block throws of varying measured trajectories $\tau$ (from real dataset ContactNets) used to estimate unknown inter-body block signed distance functions. Contact-based data collected from block trajectories contain varying amounts of information about parameters, indicated by color concentrations across the horizontal span of parameters of interest $\theta$. Our goal is to regulate and rank contact data by its information content to quickly and efficiently converge on parameter uncertainty by populating maximal amount of total information (summation over all available information from the dataset) across all parameters.
• Figure 2: Dataset Quality Improves Performance. Here, we show the performance over three models written in ContactNets for three selected performance metrics over fixed numbers of data sizes. We observe overall improvement of our traced Fisher-ordered ranking approach across all methods evaluated across three samples per experiment of fixed data size. We observe weaker improvement over end-to-end experiments due to the degeneracy of gradients of fitting a network directly to the contact impulses. By contrast, we see stronger improvement, on average, for the ContactNets polytope and deepvertex ranked methods due to the explicit physics-based structure applied to both models, significantly improving parameter sensitivity gradients. Additionally, we show, on average, that our approach generates lower variance of trajectory and penetration errors over the sampled experiments.
• Figure 3: Experimental Design for Learning. We show two different robot demonstrations from our proposed contact-aware experimental design approach. The robot is tasked with selecting optimal initial throws that result in information-rich contact data (block trajectory measured by AprilTags). In the first throw, we indicate that improving on the initial conditions of optimal throws results in high excitations in rotation from contact, thus flipping the block over upon contact to identify inter-body signed distance functions for all vertices of the block. In the second throw, we indicate that our method excites information-rich contacts through high sliding forces on the block for proper identification of tangent contact Jacobians.
• Figure 4: Fisher Functionals Invariably Improve Learning Here, we show results of our Fisher-order ranking method using different functional $\Phi(\cdot)$ mappings: trace, minimum eigenvalue, and our info-orthogonal ranking method. All methods reveal slightly different, but similar, extractions of Fisher information for the curation of data for learning. However, we anticipate that under the presence of datasets with largely diverse informativeness of parameters, the info-orthogonal methods improve the data ranking quality under low-data regimes.
• Figure 5: Optimal Data Generation via Fisher Information Maximization. Using our algorithm, we optimize over the contact-aware Fisher information matrix to generate an initial block throw that generates expected information-rich contact data used for learning. We utilize the trace of the Fisher information as our matrix reduction operator. Our approach is compared with a normal random sampling method that samples random initial block throws and executes the throw on the real robot. The mean and variance of the normal distribution used to sample initial throws is calculated empirically from the existing ContactNets dataset. Our results show overall improvement by explicitly reasoning over the information in contact compared to random sampling, thus boosting the ContactNets DeepVertex learning process.

Theorems & Definitions (8)

• Definition 1
• Definition 2
• Proposition 1
• Definition 3
• Definition 4
• Theorem 1
• proof
• Definition 5