Learning to Absorb Corrections: Critic-Weighted Residual Flow for Vision-Based Insertion
We present ACT--Flow, an online learning framework that couples an adaptive Action Chunking Transformer (ACT) with a conditional residual flow policy.
ProofPaper ↗
Key points
- Vision-based insertion requires correcting alignment errors while retaining useful behavior learned from demonstrations.
- Following anchor-guided warmup, ACT absorbs successful executed actions through behavioral cloning, while residual targets are recomputed relative to the updated prior.
- A chunk-level critic learns from successful and failed interactions and evaluates final composed actions.
- Component ablations further support the complementary roles of behavior absorption, value weighting, and successful-action supervision.
Sources (1)
- [1]Learning to Absorb Corrections: Critic-Weighted Residual Flow for Vision-Based InsertionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:00 AM
We present ACT--Flow, an online learning framework that couples an adaptive Action Chunking Transformer (ACT) with a conditional residual flow policy.
Vision-based insertion requires correcting alignment errors while retaining useful behavior learned from demonstrations.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026KRAFTON/Raon-1-mini-preview
- Oct 20, 2026Beyond the KV cache: Training and serving i14’s Laplace 3B linear attention model on AWS Trainium2 with native PyTorch
- Oct 20, 2026JUICE: Unified Multimodal Modeling in Pretrained Representation Spaces
- Oct 20, 2026UniWALM: Unified World Action Language Model for Embodied Robot Control
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026unsloth/flint-GGUF