With the growing use of wearable tactile sensor arrays, tactile sensory outcomes (TSOs) provide valuable insight into a robot’s dynamic responses to its actions. Accurately estimating a robot’s dynamic state during physical human-robot interaction (pHRI) is essential for handling real-world uncertainty. Enabling robots to adapt effectively in pHRI remains a key challenge for embodied intelligence, making reliable dynamic state estimation during interactions critically important.

An article published in IEEE Robotics and Automation Letters introduces a new method that helps robots predict how they will feel during physical interactions with humans. The researchers propose a core idea: rather than depending on complex physics-based models or extensive historical data, they introduce an embodied learning approach. This method enables robots to learn from experience and predict tactile outcomes in real time using only current inputs, such as tactile signals and action data.

The contributions of this study include:

  • Introduces DL-EH, a deep learning–based embodied haptic model for predicting tactile sensory outcomes.
  • Proposes a label-free training method that uses historical data during training but enables real-time prediction from current inputs only.
  • Demonstrates strong generalization in real-world scenarios and validates performance through a robotic framework for pHRI.

After outlining related works and current limitations, the researchers present their proposed method.

Proposed Model: DL-EH

The authors developed a deep learning model called DL-EH (Deep Learning–Embodied Haptic model). Instead of relying on traditional methods that require manual modeling and large historical datasets and struggle with real-time performance and generalization, DL-EH predicts touch outcomes instantly, with no history needed at runtime, adapts to new interaction scenarios, and works well even with uncertainty in human-robot contact.

An application example of the tactile sensory outcome prediction in the physical human-robot interaction framework.

 

Predicting TSOs in human–robot interaction requires linking robot actions to tactile signals across space and time. While historical tactile data helps capture these action-induced patterns, using it directly can hinder real-time performance. 

Inspired by an innovative paradigm for embodied intelligence, the authors propose an embodied learning approach that uses past data only during training to learn these patterns implicitly. The DL-EH model combines an assist module that extracts patterns from tactile sequences with a forward predictor that uses only current inputs to generate TSOs. 

Overview of DL-EH for pHRI. The assist module uses tactile sequence data to help the forward predictor learn to predict TSO from current measurements.

 

After training, the system can efficiently predict tactile outcomes in real time, supported by a structured three-step training process.

Data streaming process.

 

Experiment Results

The experiments used a humanoid robot equipped with tactile sensors to collect interaction data without human intervention, enabling self-supervised learning of tactile dynamics. The DL-EH model was trained and evaluated on this data, with performance measured by prediction accuracy. 

An example of the data acquisition process.

 

Results showed that the model achieves strong performance with an efficient, compact design, benefits from optimized hyperparameters, and outperforms several baseline deep learning approaches. It also handles uncertainty in interactions, generalizes well to new scenarios, and improves further when combined with vision. Importantly, DL-EH supports real-time prediction and enhances robot control tasks—reducing delays, smoothing motion, and improving safety—demonstrating strong potential for practical human-robot interaction applications.

Future Work

To further improve DL-EH training, future work will explore robot learning strategies, such as model-based reinforcement learning, to enable more efficient and adaptive actions in real-world environments. 

This approach will also be extended to robot control, allowing actions to be guided by predicted tactile outcomes (TSOs), so robots can learn which behaviors correspond to specific sensory results. Additionally, DL-EH could be combined with multimodal models to evaluate potential future actions by predicting their TSOs, ultimately supporting more autonomous and intelligent robot planning.

Interested in learning more about Robotic and Human-Machine Interactions? We have thousands of articles related to these industries and more! Also learn more about our eBooks and eLearning collections on cutting edge technologies.

Interested in acquiring full-text access to this collection for your entire organization? Request a free demo and trial subscription for your organization.