[NeurIPS 2026] VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
-
Updated
Oct 26, 2026 - Python
[NeurIPS 2026] VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
Code for Scalable Offline Model-Based RL with Action chunking
multi-robot serving engine for cloud robot foundation models
SequenceLink serves robot policies from GPUs anywhere: the policy runs on a GPU, and the robot connects to it over the internet. Works with openpi, LeRobot, GR00T and Octo, and with any policy written as a class. pip install sequence-ai
Official simulation reproduction for SpeedTuning (ICRA 2025).
End-to-end implementation of Action Chunking Transformers (ACT) for imitation learning in robot manipulation tasks. Trained and evaluated in MuJoCo simulation on a pick-and-place task using the ROBOTIS FFW humanoid robot.
Learning pick-and-place with ACT on a low-cost CPU robot: demonstrations, asynchronous deployment, controller filtering and real-world evaluation.
Four VLA action representations compared on multimodal demonstrations: bins, regression, diffusion, flow. Regression drives into the obstacle it should route around; flow keeps multimodality at 90x diffusion's control rate.
A list of Action chunking RL papers (continually updated)
GRACE: Geometric Memory for Adaptive Action Chunking. Method core from the paper supplementary material.
SmolVLA 多任务适配与异步动作执行研究 — 数据效率与语言消融、RTC 异步执行
Evidence-aware evaluation infrastructure for embodied AI
ABK-VLA: Task-ID-Free Continual Visuomotor Policy Learning via Autonomous B-Spline Knot Compression
World-model assisted VLA-lite prototype for offline robot control with action-chunk reranking.
Controlled XR-1 VLA study of action-prefix conditioning and asynchronous temporal alignment.
FoldingACT: multimodal (vision + tactile) imitation-learning policy for bimanual dexterous paper folding, IROS 2026 Robotic Origami Challenge
How should ACT execute its action chunks under observation latency and a moved target? A LeRobot study in ALOHA simulation: 21,600 paired episodes on two checkpoints.
A CPU-first end-to-end imitation learning lab: collect demos, train BC/action chunks, evaluate closed loop, write reports.
GR00T-style asynchronous EEF action-chunk execution for G1/Inspire, with a hardware-free reproducible demo.
Reproducing Action Chunking Transformer (ACT) on a low-cost SO-101 arm.
To associate your repository with the action-chunking topic, visit your repo's landing page and select "manage topics."