docs/DOWNSTREAM_GWAM_INSTRUCTIONS.md

Downstream Graph-WAM Instructions

The next student should treat this dataset as graph-augmented robot trajectory data for DreamZero-style joint vision+action flow matching.

Do not assume graph conditioning is the main path

The graph is not assumed to be the main conditioning input. It can be used in several ways, in this recommended order:

1. Structured modality/state stream:

history RGB + robot state + language + graph state -> joint future vision/action model

2. Auxiliary ΔG target:

predict future video/action plus relation/event changes

3. Joint denoised graph modality:

flow match [future video latent, future action, future graph state]

4. Conditioning ablation only:

same model with graph tokens as extra conditioning

The dataset intentionally supports all of these.

Default chunk builder

For chunk starting at t:

context_rgb = frames[t-L:t]
context_state = robot_state[t-L:t]
context_graph = graph_tensors[t-L:t]
target_rgb = frames[t:t+H]
target_action = action[t:t+H]
target_graph_delta = events_or_transition_masks[t:t+H]
language = episode.language_instruction

Required ablations

• no graph;
• graph shuffled across episodes;
• stale graph from earlier timestep;
• state-tier only;
• prior-tier only;
• graph as auxiliary target only;
• graph as conditioning stream;
• graph as joint denoised modality;
• right-only vs left-only vs right+wrist vs all three views.

Metrics

Use both visual/action metrics and graph-state metrics:

• action MSE / control accuracy / rollout success;
• video reconstruction/prediction metrics;
• relation event F1;
• contact/support/articulation ΔG accuracy;
• target-object visibility/ROI metrics;
• corruption sensitivity.