When Sensing Varies with Contexts: Context Probing for Tactile Few-Shot Class-Incremental Learning

Yifeng Lin, Aiping Huang, Wenxi Liu, Si Wu, Tiesong Zhao, Zechao Li, Zheng-Jun ZhaarXiv 2026 · arXiv

TL;DR

Tactile few-shot class-incremental learning (FSCIL) suffers because the exact same material can produce wildly different signals depending on how it is touched (e.g., varying force, speed, or sensor types). This paper introduces CoP-FSCIL, a context-aware framework that actively probes and suppresses these context-sensitive variations to extract robust, material-intrinsic features. By separating what a material is from how it was sensed, the method sets a new state-of-the-art on both tactile and audio benchmarks.

What Problem Does This Paper Address?

Few-Shot Class-Incremental Learning (FSCIL) requires AI models to learn new categories over time from just a few examples without forgetting what they previously learned. While this works reasonably well for images, it breaks down completely in robotic tactile sensing.

Unlike a standard camera image, a tactile signal is deeply entangled with its acquisition context. If a robotic finger scrapes a piece of velvet quickly with high force, the resulting data looks entirely different than if it scraped it slowly with a light touch. When a model tries to learn a new material from only a few samples, it easily falls into two traps:

  1. Context Overfitting: The model mistakenly learns the speed or force of the trial as a defining feature of the material itself.
  2. Context Variation Failure: When encountering a familiar material under a brand-new context (e.g., a new scanning speed), the model fails to recognize it.

This paper tackles the challenge of eliminating this "context-induced prototype bias" without requiring explicit, manual annotations of the context conditions.

Acquisition contexts can cause context variation failure and context overfitting during tactile learning.

Methods

The authors propose Context-Probing Few-Shot Class-Incremental Learning (CoP-FSCIL), which consists of three interconnected components:

  1. Context-Probing Intervention (CPI): To understand how context alters data, the model actively generates simulated variations of the input signal. It applies low-level perturbations (like global energy biases, frequency tilts, and temporal shifts) and mixes in statistical profiles from a "Context Statistic Bank" containing diverse, previously observed sensing styles.
  2. Probe-Conditioned Quotient Adapter (PCQA): Once the model has the original sample and its perturbed variants, it uses a lightweight network to predict a series of Householder reflections. These reflections form a sample-adaptive orthogonal coordinate transformation. Crucially, this math allows the model to cleanly separate the embedding space into two blocks: a probe-invariant block (the true material properties) and a probe-sensitive residual block (the context noise). The context noise is then zeroed out.
  3. Probe-Stability Prototype Calibration (PSPC): During incremental learning sessions, the model evaluates how much a new sample's embedding fluctuates under simulated probing. If a sample changes drastically, it is deemed "unstable" (highly context-contaminated) and down-weighted when calculating the final class prototype.
The overall architecture of CoP-FSCIL, highlighting the CPI, PCQA, and PSPC modules.

Datasets

The framework was evaluated on two primary numerical-tactile datasets and two extended audio datasets:

  • HapTex: A tactile dataset containing 100 fine-grained material classes (fabrics like velvet, leather, silk) with time-series measurements of friction force, normal force, displacement, and velocity.
  • LMT108: A multimodal tactile dataset featuring 120 material classes recorded via a handheld pen-like sensor containing a triaxial accelerometer, force resistors, and friction sensors.
  • LS-100 & NS-100: Two audio benchmarks derived from LibriSpeech (speaker identity) and NSynth (musical instruments). Audio signals share similar spectro-temporal properties with tactile data, where task-irrelevant factors like volume, pitch, or speaker emotion act exactly like "acquisition contexts."

Key Insights and Ideas

  • Orthogonal Separation over Erasure: Instead of trying to forcefully erase context information via standard neural layers—which can inadvertently destroy semantic data—CoP-FSCIL uses orthogonal transformations. This preserves the underlying geometric distances and inner products of the feature space while neatly isolating the noise.
  • Stability Equals Reliability: In few-shot learning, not all data points are created equal. Measuring how much an embedding "wiggles" when subjected to artificial context shifts provides an elegant, unsupervised metric for how reliable that sample is for prototype generation.
  • Tactile to Audio Parallelism: A fascinating takeaway is that "acquisition context" in touch behaves identically to "acoustic environment/speaker style" in audio. The exact same architecture designed for robotic fingers successfully prevents acoustic overfitting in speech and music models.

Key Results

CoP-FSCIL was benchmarked against several state-of-the-art FSCIL methods (including OrCo, PITS-SC, and PA-PCT) using Average Accuracy (AA) across all incremental steps and Performance Drop (PD) from the first to last session.

  • HapTex Benchmark: CoP-FSCIL achieved an Average Accuracy of 84.00%, outperforming the closest baseline (PA-PCT) by 2.52 percentage points.
  • LMT108 Benchmark: On this highly challenging handheld dataset, it boosted Average Accuracy to 37.34% (a 3.39% improvement over the runner-up).
  • Audio Generalization: On the LS-100 and NS-100 audio datasets, CoP-FSCIL maintained the top spot, proving its algorithmic generality.
  • Efficiency: Because the PCQA module is applied after the heavy feature extraction backbone as a lightweight adapter, it introduces negligible computational overhead in terms of FLOPs and parameters compared to standard baselines.

Limitations and Future Work

While powerful, CoP-FSCIL relies on generating a relatively large number of context probes ($N_{ps} = 100$ for HapTex) to accurately calculate a sample's stability. This multi-probe requirement increases the time needed for incremental updates during deployment.

The authors note that future research will focus on adaptive probe selection (generating only the most informative perturbations rather than a broad sweep) and incorporating physics-aware probe generation to better mimic real-world sensor mechanics.

Summary updated

Abstract

Few-shot class-incremental learning (FSCIL) aims to recognize novel classes from only a few labeled samples while retaining previously learned knowledge. Although recent FSCIL methods have achieved substantial progress on visual benchmarks, they remain limited in tactile sensing, where the same material may produce markedly different observations under different acquisition contexts, such as sensing devices, contact states, scanning trajectories, and interaction conditions. In tactile FSCIL, the challenges of few-shot learning and class-incremental learning are further amplified by acquisition context: the limited support samples may not only be scarce, but also carry context-induced biases. Once the resulting biased prototypes are inserted into the classifier, they may affect the decision boundaries in subsequent sessions. To address this problem, we propose Context-Probing Few-Shot Class-Incremental Learning (CoP-FSCIL), a context-aware framework for tactile FSCIL. CoP-FSCIL first employs Context-Probing Intervention (CPI) to diagnose local context-sensitive variations in tactile representations. It then introduces a Probe-Conditioned Quotient Adapter (PCQA) to suppress context-sensitive components identified by the probes. Finally, Probe-Stability Prototype Calibration (PSPC) estimates support sample reliability from probe-induced embedding fluctuations and calibrates stochastic prototypes accordingly. Experiments on HapTex and LMT108 show that CoP-FSCIL consistently outperforms representative FSCIL baselines, and extended experiments on audio FSCIL further demonstrate the generality of the proposed context probing mechanism. The source code is currently being prepared and will be released soon.
BibTeX
@misc{lin2026when,
  author = {Yifeng Lin and Aiping Huang and Wenxi Liu and Si Wu and Tiesong Zhao and Zechao Li and Zheng-Jun Zha},
  title = {{When Sensing Varies with Contexts: Context Probing for Tactile Few-Shot Class-Incremental Learning}},
  year = {2026},
  eprint = {2603.25115},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.25115},
}
Figure