First do not fall: learning to exploit a wall with a damaged humanoid robot

TL;DR

This work tackles the challenge of preventing falls in damaged humanoid robots by leveraging a fast wall-contact reflex. It introduces D-Reflex, a neural-network classifier that, given the robot posture $q$ and wall configuration $(d,\alpha)$, selects a wall-contact location $(x^*,y^*)$ which is enforced through a whole-body controller to stabilize the robot. The approach is trained in simulation on a wide range of simulated damage scenarios and evaluated against baselines, showing about $76\%$ success on avoidable falls and demonstrating robustness to different damage types and wall/friction conditions; real-robot tests corroborate fast execution times (~$16$–$21$ ms). The method does not rely on precise damage-model identification during the fall, instead providing a rapid, generalizable first-response mechanism that enables damaged robots to maintain balance and continue missions, while model updates and recovery strategies follow post-stabilization. This reflex-based strategy is particularly relevant for indoors-dominated, high-risk environments where walls are commonly available as support during emergencies.

Abstract

Humanoid robots could replace humans in hazardous situations but most of such situations are equally dangerous for them, which means that they have a high chance of being damaged and falling. We hypothesize that humanoid robots would be mostly used in buildings, which makes them likely to be close to a wall. To avoid a fall, they can therefore lean on the closest wall, as a human would do, provided that they find in a few milliseconds where to put the hand(s). This article introduces a method, called D-Reflex, that learns a neural network that chooses this contact position given the wall orientation, the wall distance, and the posture of the robot. This contact position is then used by a whole-body controller to reach a stable posture. We show that D-Reflex allows a simulated TALOS robot (1.75m, 100kg, 30 degrees of freedom) to avoid more than 75% of the avoidable falls and can work on the real robot.

Figures (4)

• Figure 1: (top) A humanoid robot detects a fault on one of its legs. If it does nothing, it falls; but it can recover its stability by putting its "hand" on the wall at the appropriate location (depending on its posture, the wall distance, and the wall orientation). (bottom) The percentage of avoidable falls (see Sec. \ref{['sec:avoidable']} for the definition of "avoidable") decreases when the delay taken between damage and reflex increases; the robot has only a few milliseconds to react with a high success rate.
• Figure 2: Overview of D-Reflex. In the first step (Data Collection), we sample different situations: wall configurations (distance and orientation), robot postures, and damage combinations, and simulate the behavior of the robot for each possible point on a 21$\times$ 21 grid on the wall (441 simulations for each situation). In the second step, we train a neural network classifier to predict the success (avoiding the fall) for each point of the grid. In the third step, we query the trained neural network for each point of the grid to select the best contact position, and we set it as a contact constraint in the whole-body controller.
• Figure 3: Examples of damaged situations and corresponding contact or falling postures using D-Reflex on the TALOS talos_ref robot in simulation and on the real robot, the video https://youtu.be/hbuWr-ZNAtg shows different examples. We also show the contact map estimated by our neural network and the true contact map measured but unknown during training. We distinguished two kinds of failures: when the predicted contact map is too smooth and optimistic compared to the truth, and when there is no successful contact point, i.e., the fall is unavoidable by seeking a contact on the wall.
• Figure 4: (a) Success rate considering only avoidable situations. The box plots show the median and quartiles, the bullets being the values of the evaluations. For each variant, the learning algorithm was run 20 times on different splits of the dataset. All methods are significantly different from one another using a t-test with Bonferroni correction (p-value $\le0.001$) except D-Reflex and the J Addition. (b) Percentage of avoidable situations and situations avoided using D-Reflex, with a model trained for the friction $1$, depending on the wall friction coefficient, plus examples of contact maps.