v1.2.1 - GraphDDPM

What's Changed

• 130 add rotation to object observartions as node features by @caio-freitas in #131
• 123 add node id to node features embedding by @caio-freitas in #134

Full Changelog: v1.2.0...v1.2.1

v1.2.0

What's Changed

• fix last_naction to account for different #objects by @caio-freitas in #122
• 118 prepare script for evaluation of trajectory quality by @caio-freitas in #124
• 127 fix pos in graph dataset by @caio-freitas in #128
• 108 review masked values for nodes in argd policy by @caio-freitas in #125

Full Changelog: v1.1.9...v1.2.0

Joint Values as Node Features

What's Changed

• 107 assure correct usage of the egnn model by @caio-freitas in #117
• 119 add joint values to observation again by @caio-freitas in #120

Full Changelog: v1.1.8...v1.1.9

Graph Transport + Noise Addition to Last Actions

What's Changed

• Remove Positional Encoding by @caio-freitas in #95
• Fix BC Policy with E_GNN by @caio-freitas in #96
• add validation loss to ARGD Policy by @caio-freitas in #101
• DDPM Policy with EGNN model by @caio-freitas in #98
• Robomimic policies by @caio-freitas in #75
• Normalize Data and use 6D Rotation by @caio-freitas in #105
• 100 add transport graph environment by @caio-freitas in #110
• 102 start ddpm policy from last action by @caio-freitas in #111
• use global learning rate scheduler step in ARGD and GraphDDPM policies by @caio-freitas in #115

Full Changelog: v1.1.3...v1.1.8

BC-RNN Robomimic Lowdim Policy

What's Changed

• Robomimic policies by @caio-freitas in #75

Full Changelog: v1.1.3...v1.1.5

EGNN Condition Encoder w/ Joint Velocities

What's Changed

• setup square_graph task by @caio-freitas in #89
• add GraphConditionEncoder by @caio-freitas in #90
• 79 investigate discretizing gripper control or data by @caio-freitas in #87
• add sweep script based on train by @caio-freitas in #93
• Improve Graph Representation by @caio-freitas in #94

Full Changelog: v1.0.4...v1.1.3

EGNN Condition Encoder

What's Changed

• setup square_graph task by @caio-freitas in #89
• add GraphConditionEncoder by @caio-freitas in #90

Full Changelog: v1.0.3...v1.1.1

Equivariant IfO Autoregressive Graph Diffusion Model

1. Graph structure modified to contain task-space positional information in the g.pos attributes (position and quaternion orientation).
2. Environment (robomimic_graph_wrapper) updated to match with 1.
3. E(N) Equivariant Graph Neural Networks from Satorras et al. layers (E_GCN) used instead of old message passing layers.
4. All horizons (at least initially) set to 1, to deal with the node positions
5. Loss function updated to separately account for the positions and joint values, with weighting coefficients.

Full Changelog: v1.0.2...v1.0.3

IfO Autoregressive Graph Diffusion Model

• Using task-space as observation (with joint values set to 0)
• Addition of lr scheduling
• Addition of steps in the graph dataset to avoid idling (using 1 as step in the end of the episode)
• Optimizing w.r.t. all node features (task-joint-space graph dataset)

Full Changelog: v1.0.0...v1.0.2

First Autoregressive Graph Diffusion Model

This release contains the first Autoregressive Graph Diffusion Model for behavioral cloning to be fully benchmarked. Using only joint values as node features, and appending the observation and prediction horizon features to the node features. The forward and reverse absorption process is made node by node as in the Autoregressive Diffusion Model for Graph Generation from Kong et al.