Event
TTT and Human Demonstration Training Did Not Meet Expectations for VLA Models
1 reports · 1 sources
Qiangong Zhou and other authors released a technical report integrating test-time training (TTT) into InternVLA-A1 and Pi0.5, training with paired human demonstrations and robot trajectories to investigate action generation conditioned on human demonstrations. The report states that at the current scale, adding TTT and paired-data training did not yield consistent demonstration-based target selection; shortening sequences, full-trajectory gradient connectivity, and contrastive regularization also did not noticeably improve the model's response to demonstrations. When demonstrations were reduced to single-frame cues, real-robot execution still preferred the cup at a fixed position.
Report timeline
- TTT in VLA Models: Exploring Training with Human Demonstrations in InternVLA-A1 and Pi0.5
The study integrates Test-Time Training (TTT) into InternVLA-A1 and Pi0.5, training on paired demonstrations and robot trajectories to explore action generation conditioned on human demonstrations. At the current scale, adding TTT and training on paired data did not produce consistent target selection based on demonstrations, and shorter sequences, full-trajectory gradient connectivity, and contrastive regularization also did not clearly improve responsiveness to demonstrations. When demonstrations were reduced to single-frame cues, real-robot execution still preferred cups at fixed positions.
arXiv cs.RO ·