The paper introduces CRAFT, a method for improving compositional generalization in vision‑language‑action (VLA) models. It addresses the issue where models rely on visual shortcuts during fine‑tuning, leading them to execute demonstrated skill combinations that match observations rather than the instructed ones. By training with counterfactual instruction–observation pairs and transferring supervision through reusable skill representations, CRAFT enhances success on unseen skill combinations while preserving performance on demonstrated ones across multiple VLA models and benchmarks.
By Taegeun Yang, Youngju Na, Yoonki Cho, Sung-Eui Yoon
arXiv:2609.37359v1 Announce Type: cross
Abstract: Coding agents can now write, run, and debug programs with little human help. Robot tasks, however, are usually specified by a sentence that leaves ou...
By Yifan Kang, Zihan Wang, Zhiwen Fan, Bangya Liu
DexPIE is a post‑training framework that improves dexterous manipulation policies using real‑world experience. It introduces a dexterous‑hand‑adapted intervention system and multi‑stage DAgger‑style data collection to enhance exploration, aligns training and inference to reduce distribution shift, and conditions the policy on a continuous optimality indicator for fine‑grained data quality use. In three real‑world tasks, DexPIE boosts success rates by 37.3% over a demonstration‑based baseline, outperforming all other methods and showing stronger robustness.
By Ruizhe Liao, Wenrui Chen, Liangji Zeng, Haoran Lin, Fan Yang, Kailun Yang, Yaonan Wang
arXiv:2607. 01754v1 Announce Type: new Abstract: On-policy exploration is a crucial component for training robust Vision-Language Navigation agents, as it exposes the policy to a broader state distribution.
By Sung June Kim, Sangpil Kim, Honglak Lee
The paper introduces Spatial Language Modeling, a framework that unifies action generation and state prediction using a shared vocabulary of discrete coordinates and semantic tokens. An autoregressive Transformer learns both tasks through a common next‑token objective, trained first with random‑play transition pretraining and then on expert demonstrations. Experiments on the Push‑T task in simulation and on a real robot show that the model matches simulation performance and outperforms baseline policies in task success and target coverage, while also accurately predicting scene states from action trajectories.
By Minye Wu, Zehao Wang, Tinne Tuytelaars
SynthDemo‑RL introduces a teacher‑student framework that uses an automated teacher to generate successful manipulation trajectories from simulator‑privileged state, which are then distilled into a Vision‑Language‑Action (VLA) student via supervised fine‑tuning. The student is further refined with PPO using binary task‑success rewards. On the LIBERO‑PRO benchmark, SynthDemo‑RL rescues all 27 previously unsolvable tasks and achieves near‑perfect success rates, matching performance that would otherwise require human demonstrations.
By Hiroaki Kingetsu, Hiroaki Kurihara, Kaoru Yokoo, Kenji Fukumizu, Manohar Kaul