Long-Horizon Visual Planning with Goal-Conditioned Hierarchical Predictors
Supplementary Material
Video Results
Pick&Place
Ground Truth
GCP-tree
GCP-sequential
Deep Voxel Flow
(Liu'17)
CIGAN
(Wang'19)
Reconstructions on the Pick&Place data. The sequences contain 80 64x64 frames.
GCP-tree
GCP-tree
GCP-sequential
GCP-sequential
Prior samples from GCP on the Pick&Place data. Each column represents different conditioning information. The sequences contain 80 64x64 frames.
Human 3.6 Million
Ground Truth
GCP-tree
GCP-sequential
Deep Voxel Flow
(Liu'19)
CIGAN
(Wang'19)
Reconstructions on the H36 dataset. The sequences contain 500 64x64 frames.
GCP-tree
GCP-tree
GCP-sequential
GCP-sequential
Prior samples from GCP on the Human 3.6M dataset. Each column represents different conditioning information. The sequences contain 500 64x64 frames.
9 Rooms
Ground Truth
GCP-tree
GCP-sequential
Deep Voxel Flow (Liu'17)
CIGAN (Wang'19)
Reconstructions on the 9-room data. The sequences contain 100 32x32 frames.
GCP-tree
GCP-tree
GCP-sequential
GCP-sequential
Prior samples from GCP on the 9-room data. Each column represents different conditioning information. The sequences contain 100 32x32 frames.
25 Rooms
Ground Truth
GCP-tree
GCP-sequential
Deep Voxel Flow (Liu'17)
CIGAN (Wang'19)
Reconstructions on the 25-room data. The sequences contain 200 32x32 frames.
GCP-tree
GCP-tree
GCP-sequential
GCP-sequential
Prior samples from GCP on the 25-room data. Each column represents different conditioning information. The sequences contain 200 32x32 frames.
Visual Control on Navigation
Comparison of visual planning & control approaches. Execution traces of Visual Foresight (left), GCP-tree with non-hierarchical planning (middle) and GCP-tree with hierarchical planning (right) on two 25-room navigation tasks. Visualized are start and goal observation for all approaches as well as predicted subgoals for hierarchical planning.