Event
inr2action Encodes Robot Memory with Fixed-Size Weight Deltas
1 reports · 1 sources
Leann Alhashishi, Christos Margadji, and Sebastian William Pattinson propose a parameter-space memory framework, inr2action, which fits a neural implicit representation for each robot manipulation episode starting from a shared meta-learned initialization, using its weight delta as the memory; the size does not vary with episode length and remains fixed from one hundred frames to one hundred thousand frames. On the RoboMME benchmark, the language model exceeds chance level on all question types, with the decoder achieving an average R² of 0.51; replacing the weight delta with an equal amount of noise causes an 11.4-point performance drop, while weights from random starting points cannot be decoded, on which basis the authors claim that the meta-learned initialization is necessary for embedding episode content.
Report timeline
- inr2action: Encoding Robot Episodes as Neural Implicit Representations for Recall and Action
The paper proposes the inr2action framework, which fits a neural implicit representation to each robot manipulation episode starting from a shared meta-learned initialization; the resulting weight delta serves as memory, with a size that does not vary with episode length and stays fixed from one hundred frames to one hundred thousand frames.
arXiv cs.RO ·