PARA

Testing how PARA's pixel-aligned formulation generalizes to out-of-distribution object positions and camera viewpoints. Comparing robustness and data efficiency against global-regression baselines (ACT, DINO-VLA, InternVL).

• Does PARA generalize better to unseen object positions?
• How does performance degrade under viewpoint shift?
• Is PARA more data-efficient than global-regression baselines?

Comparing PARA's pixel-aligned regression head vs. global regression on top of a video generation backbone (UVA). Testing the hypothesis that PARA is more data-efficient for learning joint video-action policies.

• Does PARA head outperform global regression on video backbone?
• How does video conditioning improve action prediction?
• What is the data efficiency gain from pixel-aligned prediction?

Pretraining PARA on the large-scale DROID dataset (100K+ trajectories) to test whether pixel-aligned prediction benefits from diverse cross-embodiment data.

• Does DROID pretraining improve downstream LIBERO performance?
• How does pretraining scale with dataset size?
• Can PARA transfer across embodiments via pixel alignment?

Deploying PARA on a real Franka Panda arm to validate sim-to-real transfer and real-world pixel-aligned action prediction.

• Does PARA's pixel-aligned prediction transfer to real hardware?
• How sensitive is performance to camera calibration?
• Can real-world PARA match or exceed sim baselines?