Goal-Conditioned Hierarchical Prediction

Supplementary Material

Planning

Plan executions. Left: our model, middle: GCBC, right: Ebert et al. 2018.

Pick&Place

Reconstructions on the Pick&Place environment. Top to bottom: target sequence, GCP-tree, GCP-sequential, Deep Voxel Flow, Causal InfoGan. The sequences contain 80 64x64 frames.

Prior samples from GCP on the Pick&Place environment. Each column represents different conditioning information. The two top samples are from GCP-tree and the two bottom samples are from GCP-sequential. The sequences contain 80 64x64 frames.

Human 3.6 Million

Reconstructions on the H36 dataset. Top to bottom: target sequence, GCP-tree, GCP-sequential, Deep Voxel Flow, Causal InfoGan. The sequences contain 500 64x64 frames.

Prior samples from GCP on the Human 3.6M dataset. Each column represents different conditioning information. The two top samples are from GCP-tree and the two bottom samples are from GCP-sequential. The sequences contain 500 64x64 frames.

3x3 Maze

Reconstructions on the 3x3 Maze environment. Top to bottom: target sequence, GCP-tree, GCP-sequential, Deep Voxel Flow, Causal InfoGan. The sequences contain 100 16x16 frames.

Prior samples from GCP on the 3x3 Maze environment. Each column represents different conditioning information. The two top samples are from GCP-tree and the two bottom samples are from GCP-sequential. The sequences contain 100 16x16 frames.

10x10 Maze

Prior samples from GCP on the 10x10 Maze environment. Each column represents different conditioning information. The two top samples are from GCP-tree and the two bottom samples are from GCP-sequential. The sequences contain 1000 16x16 frames.

Page updated

Google Sites

Report abuse