OMRON SINIC X Corporation (HQ: Bunkyo-ku, Tokyo; President and CEO: Masaki Suwa; hereinafter “OSX”) is pleased to announce that two of its latest research papers have been accepted for The Conference on Robot Learning　(hereinafter “CoRL 2026”).

CoRL 2026 is one of the top-tier international conferences in Robot Learning. In 2026, 687 papers (approximately 33%) were accepted out of 2,094 submissions, and the conference will be held from November 9 to November 11, 2026, in Austin, Texas, United States (local time).

The research papers to be presented by OSX are as follows.

Papers Accepted at CoRL 2026

■ CLAP: Direct VLM-to-VLA Adaptation via Language-Action Grounding

Yuri Ishitoya (Ochanomizu University / OSX Intern), Jeremy Siburian (The University of Tokyo / OSX Intern), Masashi Hamaya (OSX), Haruto Suzuki (OSX Intern), Kuniaki Saito (OSX), Toshihiko Fukushima (OSX), Cristian Camilo Beltran-Hernandez (OSX), Mai Nishimura (OSX)

Vision-language models (VLMs), which jointly understand images and language, have advanced rapidly, and the latent capability of VLMs alone to control robots is drawing attention, as seen in large general-purpose models that operate robots directly. In contrast, existing vision-language-action models (VLAs) require dedicated action output modules, architectural modifications, or additional training on large-scale robot data, which substantially degrades the capabilities of the original VLM. This makes it difficult to analyze which of the VLM’s inherent capabilities contribute to robot control, and slows the adoption of VLM advances in VLA research.
We propose CLAP (Causal Language-Action Prediction), a method that directly adapts any pretrained VLM to robot control (VLA) without modifying its architecture. Training a VLM to predict robot actions directly as numeric token sequences pushes its outputs far from the pretrained language distribution and impairs its original capabilities. During training, CLAP automatically generates natural language action descriptions, such as “move forward 3 cm, tilt right 15 degrees,” and places them before the numeric action tokens. This encourages the model to stay closer to its pretrained distribution, converting the VLM into a VLA while mitigating the loss of its capabilities.
On the LIBERO robot manipulation benchmark, CLAP-2B achieves a 90.8% success rate with only 4 to 5 hours of post-training, outperforming VLA-0 by 14.9 points. CLAP also improves performance on LIBERO-PRO, a robustness evaluation with instruction, object, and spatial perturbations, and we confirm its effectiveness on a real robot arm. We release VLA models at 0.8B, 2B, and 4B scales, and expect this work to serve as a research foundation where any new VLM can immediately become a starting point for VLA research.

https://arxiv.org/abs/2607.08974
https://omron-sinicx.github.io/clap/

■ Online Robot Manipulation Error Detection and Description

Yung-Shun Chan (National Taiwan University), Kuniaki Saito (OSX), Cristian Camilo Beltran-Hernandez (OSX), Jeremy Siburian (The University of Tokyo / OSX Intern), Masashi Hamaya (OSX), Chun-Yi Lee (National Taiwan University)

Detecting and describing failure events during robot manipulation is critical for safe autonomous deployment, yet existing methods either reduce failures to per-frame binary scores or require the complete episode. We formulate robot failure detection as an online video temporal grounding problem, where the model processes streaming video and outputs both the temporal boundaries and a natural-language description of each failure event as it unfolds.
Our framework, STREAM-FAIL, operates chunk-by-chunk with a memory bank over previously observed frames. To bridge the domain gap between general video understanding and manipulation reasoning, we design an automated pipeline that converts robot trajectory datasets into temporal grounding annotations for pre-training, which yields a 73% relative improvement in temporal localization.
In our evaluation, STREAM-FAIL identified when robot failures occurred more accurately than existing comparison methods that process video as it arrives, and generated natural-language descriptions of those failures. We release the automatically generated dataset and the human-annotated evaluation dataset, DETECT-DROID-FAILURE, to support future research on real-time robot failure detection, description, and recovery.

https://omron-sinicx.github.io/Online_Robot_Error_Detection

※The authors’ affiliations are current as of the date of writing or submission. Please be advised that the information may become outdated after that point.

For any inquiries about OSX, please contact us here.