Pre-trained LLMs can iteratively improve dynamic manipulation policies by predicting parameter deltas from a small in-context dataset of improvement examples, outperforming low-data baselines in simulation and on a real robot.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.RO 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
In-Context Iterative Policy Improvement for Dynamic Manipulation
Pre-trained LLMs can iteratively improve dynamic manipulation policies by predicting parameter deltas from a small in-context dataset of improvement examples, outperforming low-data baselines in simulation and on a real robot.