Interpretable Hierarchical Reinforcement Learning for Continuous Control

A Study of NEXUS Across Skills, Language, and Vision

Ivan Smirnov · Maryana Smirnova · Rosela Berberi · Anja Koroveshi

Technical University of Darmstadt

Greedy rollouts of the trained controllers. Each environment is shown for the flat policy and for the neural, neuro-symbolic, and symbolic skill hierarchies. Videos loop.

Cartpole
Cheetah
Walker
Hopper
Panda
Go1 flat
Go1 rough

Cartpole

Balance an inverted pole on a cart.

Flat · seed 0 · return 948

Neural · seed 0 · return 938

NeSy · seed 0 · return 482

Symbolic · seed 0 · return 319

Cheetah

Planar cheetah runs forward as fast as it can.

Flat · seed 0 · return 912

Neural · seed 0 · return 906

NeSy · seed 0 · return 907

Symbolic · seed 0 · return 865

Walker

Planar biped walks forward at a target speed.

Flat · seed 0 · return 986

Neural · seed 0 · return 966

NeSy · seed 0 · return 872

Symbolic · seed 0 · return 760

Hopper

One-legged hopper hops forward without falling.

Flat · 8× budget, hopping · seed 0 · return 633

Neural · seed 0 · return 0

NeSy · budget-matched · seed 0 · return 169

Symbolic · seed 0 · return 0

Panda

Franka Panda reaches, grasps and lifts a cube.

Flat · seed 0 · return 1164

Neural · seed 0 · return 501

NeSy · seed 0 · return 400

Symbolic · seed 0 · return 228

NeSy · 10-step commitment · seed 0 · return 88

Go1 flat

Unitree Go1 tracks a commanded linear and yaw velocity on flat ground.

Flat · seed 2 · return 11

Neural · seed 0 · return 3

NeSy · seed 0 · return 2

Symbolic · seed 0 · return 2

Go1 rough

Unitree Go1 tracks a commanded velocity over a heightfield.

Flat · seed 1 · return 10

Neural · seed 1 · return 5

NeSy · seed 1 · return 3

RGB observations

The skill actor still receives the simulator state. These clips are the extra 64×64 grayscale camera it was given in the vision study. Side-by-side clips put the state-only rollout on the left and the state-plus-camera rollout on the right.

Cartpole, in-loop camera, seed 0. 64×64 grayscale added to the state.

Cheetah, in-loop camera, seed 0. 64×64 grayscale added to the state.

Walker, in-loop camera, seed 0. 64×64 grayscale added to the state.

Walker, in-loop camera, seed 1. 64×64 grayscale added to the state.

Cartpole, seed 0. State only on the left, state plus camera on the right.

Cheetah, seed 0. State only on the left, state plus camera on the right.

Walker, seed 0. State only on the left, state plus camera on the right.

Walker. State only, seed 0, next to state plus camera, seed 1.