Science Robotics

Supplementary Materials

The PDF file includes:

  • Note S1. Data analysis of MELA learning results.
  • Note S2. Additional Materials and Methods.
  • Note S3. Expert imbalance phenomenon.
  • Fig. S1. Comparison of MELA’s learning curves using different numbers of expert networks.
  • Fig. S2. Baseline experiments of fall recovery and trotting from engineered controllers.
  • Fig. S3. Activation patterns of experts across all motor skills.
  • Fig. S4. Five representative cases showing adaptive behaviors of the MELA expert under unseen situations in simulation.
  • Fig. S5. Forward trotting velocity during the variable-speed trotting simulation.
  • Fig. S6. Heading angle and angular velocity during the steering experiment on the real robot.
  • Fig. S7. Relative target positions with respect to the robot from the user command as the input to the MELA networks during the real locomotion experiment.
  • Fig. S8. Measured torques of the front left leg during the real locomotion experiment (Fig. 5D).
  • Fig. S9. Roll and pitch angles during the real locomotion experiment.
  • Fig. S10. Target position, body orientation, and velocity during the real locomotion experiment on grass (Fig. 5E).
  • Fig. S11. Adaptive and agile locomotion behaviors on pebbles and grass demonstrated by MELA policy.
  • Fig. S12. Continuous and variable weights of all experts during the real MELA experiment (Fig. 5D).
  • Fig. S13. Continuous and variable weights of all experts during the real MELA experiment on grass (Fig. 5E).
  • Fig. S14. Four types of unseen terrains for testing the multiskill MELA policy in the simulation.
  • Fig. S15. Simulated test scenarios for evaluating the robustness of the MELA policy.
  • Fig. S16. Representative adaptive behavior from the simulated scenario of steering on spot (fig. S4A and movie S6).
  • Fig. S17. Representative adaptive behavior from the simulated scenario of steering while recovering to trotting (fig. S4B and movie S6).
  • Fig. S18. Representative adaptive behavior from the simulated scenario of tripping (fig. S4C and movie S6).
  • Fig. S19. Representative adaptive behavior from the simulated scenario of a large impact (fig. S4D and movie S6).
  • Fig. S20. Representative adaptive behavior from the simulated scenario of falling off a cliff (fig. S4E and movie S6).
  • Fig. S21. Analysis of responses from the MELA policy during the simulated scenario of a large external perturbation (fig. S4D and movie S6).
  • Fig. S22. Phenomenon of asymmetric gait and imbalanced experts from MoE.
  • Fig. S23. Illustration of the 2D phase vector for training the locomotion policy.
  • Fig. S24. Normalized power spectrum analysis of motions during the real locomotion experiment (without the DC component).
  • Fig. S25. Learning curves during the two-stage training of MELA and MoE.
  • Fig. S26. Nine distinct configurations used as the initialization for training fall recovery policies in simulation.
  • Fig. S27. Setting of the target location for training MELA policies in simulation.
  • Table S1. Distribution matrix of expert specializations over motor skills.
  • Table S2. Specification of the Jueying quadruped robot.
  • Table S3. Detailed descriptions of the individual reward terms.
  • Table S4. Weights of the reward terms for different tasks.
  • Table S5. Selection of state inputs for different tasks and neural networks.
  • Table S6. Proportional-derivative parameters for the joint-level PD controller.
  • Table S7. Hyperparameters for SAC.
  • References (5464)

Download PDF

Other Supplementary Material for this manuscript includes the following:

  • Movie S1 (.mp4 format). Experiments of fall recovery.
  • Movie S2 (.mp4 format). Experiments of outdoor fall recovery and compliant interactions.
  • Movie S3 (.mp4 format). Experiments of trotting.
  • Movie S4 (.mp4 format). Experiments of adaptive locomotion and behaviors.
  • Movie S5 (.mp4 format). Experiments of outdoor fall-resilient locomotion on irregular terrains.
  • Movie S6 (.mp4 format). Simulation of representative adaptive behaviors from the multiskill MELA expert.
  • Movie S7 (.mp4 format). Simulation of extensive scenarios and crash tests.
  • Movie S8 (.mp4 format). Baseline experiments of fall recovery and trotting from default controllers.

Files in this Data Supplement: