Predicting Falls of a Humanoid Robot through Machine Learning

AAAI 2010

TL;DR

Humanoid robots are prone to falling, and predicting these falls is crucial for activating fall-control mechanisms. This paper introduces a machine learning approach to predict imminent falls by carefully engineering training data to balance the trade-off between early prediction (high lead time) and avoiding false alarms (low false positive rate).

What Problem Does This Paper Address?

Humanoid robots, like humans, need to maintain balance during locomotion. However, unlike humans, robots are particularly susceptible to falls due to their complex dynamics, susceptibility to external disturbances (like pushes or slippery surfaces), and the inherent limitations of their balance control systems. While simple analytical models can predict falls in idealized scenarios, they fail to capture the intricate, often non-monotonic behavior of large, high-degree-of-freedom robots under realistic conditions. This makes it challenging to develop a reliable system that can predict an impending fall with enough lead time to switch to a fall-control mode, without generating too many false alarms that disrupt normal operation.

Methods

The core of the paper is a supervised machine learning approach using decision lists (a type of rule-based classifier) to predict falls. The novelty lies not in the classifier itself, but in the data engineering and parameter tuning to achieve the desired prediction characteristics.

Key techniques include:

  • Trajectory Generation: Simulating an Asimo-like robot in Webots, applying external forces (pushes) to induce falls, and recording the resulting states.
  • Data Labeling: Classifying robot states into "balanced," "falling," and "fallen" based on physical criteria (e.g., CoM height, contact with ground).
  • Feature Engineering: Creating a 16-dimensional feature vector representing the robot's state, including CoM displacement, momenta, angular momenta, and foot contact modes.
  • Parameter Tuning for Trade-off Control:
    • T+ (Time Offset): This parameter is used during training data preparation. It involves discarding early states from "falling" trajectories, as these are harder to distinguish from "balanced" states and contribute to false positives. Increasing T+ reduces false positives but also decreases lead time.
  • This (History Window): This parameter is applied after learning. It introduces a temporal window for predictions, meaning a fall is only predicted if the classifier has consistently predicted "falling" for a certain duration. This smooths out predictions and reduces false positives, but also decreases lead time.

Datasets

The paper uses a simulated dataset generated from the Webots robotics simulator. This dataset consists of 1000 trajectories, each initiated by an external force applied to an Asimo-like robot. The authors note that while simulation results may not perfectly transfer to real robots, they establish machine learning as a viable strategy and suggest that a simulation-learned model could initialize on-line learning on a real robot.

Key Insights and Ideas

  • The Impossibility of Simple Thresholds: The paper demonstrates with Figure 2 that there isn't a simple force magnitude threshold below which falls are always avoided and above which they always occur. This non-monotonic behavior makes analytical modeling extremely difficult.
  • The Fundamental Trade-off: The most critical insight is the inherent conflict between False Positive Rate (FPR) and Lead Time (Tlead). A predictor that is too sensitive to predict falls early will generate many false alarms, while a predictor that is too conservative will miss crucial prediction windows.
  • Data Engineering as the Key: The authors argue that effective fall prediction relies heavily on how the training data is prepared and how the learning algorithm's output is post-processed, rather than solely on the choice of the learning algorithm itself.
  • The Power of Rule-Based Systems: Decision lists proved effective, suggesting that interpretable, rule-based models can capture the complex dynamics sufficiently well for this task.
  • Foot Contact Modes Matter: Separating the learning process based on foot contact modes (e.g., CoP position relative to feet) leads to more regular decision boundaries and better predictions.

Key Results

  • The proposed machine learning approach, with careful parameter tuning (T+ and This), can significantly improve upon manually designed solutions for fall prediction.
  • The paper presents three representative fall predictors (FPL1, FPL2, FPL3) illustrating different trade-offs:
  • FPL1: High Tlead (0.90s), High FPR (0.46).
  • FPL2: Balanced performance, low FPR (0.06), reasonable Tlead (0.76s).
  • FPL3: Near-zero FPR (0.005), but significantly lower Tlead (0.43s).
  • These learned predictors outperform a handcrafted solution (FPR=0.11, Tlead=0.60s).
  • The learned decision lists typically have depths of 5-10 in a 16-dimensional feature space.

Limitations and Future Work

  • Simulation vs. Real World: The authors acknowledge that performance on a real robot might differ from simulation results.
  • Batch Learning: The current approach uses batch learning, meaning data is collected and processed offline.
  • Future Work:
    • Investigate Hidden Markov Models (HMMs), which are better suited for sequential data and temporal predictions.
  • Explore on-line, incremental learning for greater autonomy, allowing robots to learn from new situations as they encounter them.
  • Conduct targeted studies on specific robot geometries, inertial properties, and controllers to further refine feature sets.

Summary updated

BibTeX
@inproceedings{2010predicting,
  title = {{Predicting Falls of a Humanoid Robot through Machine Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
}
Figure