Reproduce “Self-Improving Loops for Visual Robotic Planning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Calvin Luo, Zilai Zeng, Mingxi Jia, Yilun Du, Chen Sun
The Self-Improving Loops for Visual Robotic Planning (SILVR) framework enables in-domain visual robotic planners to continuously enhance their performance on novel tasks by iteratively learning from self-collected online experience. This approach yields up to 285% improvement in success rates on simulated MetaWorld tasks and substantial gains in real-world manipulation, outperforming baseline methods and demonstrating robustness to varying data quality.
Key insights
- Iterative finetuning of a visual planner with self-collected online experience, particularly when composed with internet-scale video priors, leads to continuous performance gains on novel robotic tasks.
- Visual planning-based self-improvement demonstrates superior sample efficiency compared to direct reinforcement learning finetuning of action-predicting policies for online adaptation.
- The framework exhibits robustness to suboptimal initial training data and varying feedback mechanisms, including VLM-based filtering or no explicit success filtering, in enabling sustained improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.