AgentRoboCritic-Muse-Glimmer-30B
AgentRoboCritic is a vision-language critic for executed robot manipulation actions. It uses the task, local intent, action command and images around execution to judge whether each action achieved its intent, categorize failures and explain the visual evidence. The released model fine-tunes Muse-Glimmer-30B.
View full model descriptionEvaluate individual executed robot manipulation actions against the stated local intent and diagnose failure mechanisms from visual observations.
View model weights (2)
Papers and technical reports
Paper / report unconfirmed
Official introductory materials are recorded; no version-specific paper or technical report has been verified.
Last checked: 2026-11-20
How to access and use
- Model files
- Public access
- Training data
- Public access
Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes
Training / fine-tuning code
Training data
Some training datasets are public; this does not establish access to the complete training mixture.
Model files
| File | Access | Weights license | Commercial use |
|---|---|---|---|
| AgentRoboCritic-Muse-Glimmer-30B (BF16) | Access | Weights license | Commercial use |
SourceVerified on · 2026-11-20
Checks before deployment
The reported evaluation covers 400 human-labeled held-out manipulation actions. The system message allows inconclusive judgments when observations are insufficient. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
Deployment and usage resources
- Runtime and frameworks
- Transformers
Architecture and inputs / outputs
- Model type
- Reward & evaluation
- Weight parameters
- 29.7766 B
- Input
- Language / RGB images
- Released
- November 20, 2026
- Input modalities · Main category: AgentRoboCritic-Muse-Glimmer-30B
- Main category: AgentRoboCritic
Scope and training
Base model
meta-models/Muse-Glimmer-30B
Pipeline task
image-text-to-text
Prediction output
JSON judgment of intent achievement, failure mechanism, grasp failure subcategory and visually grounded reason
Training data scale
12844 actions
Training data sources
https://huggingface.co/datasets/cx-cmu/AgentRoboCritic-Dataset
Lineage
Fine-tuned from: meta-models/Muse-Glimmer-30B ↗
- Base model · Pipeline task · Prediction output · Training data scale · Training data sources · Lineage · meta-models/Muse-Glimmer-30B: AgentRoboCritic-Muse-Glimmer-30B
- Training data scale · Training data sources: AgentRoboCritic-Dataset
Training data
400 held-out robot manipulation actions annotated by humans, including 264 failed actions. CC BY 4.0 covers the annotations. Images remain subject to their source benchmarks’ licenses.
400 actions · CC BY 4.0
Supervised rank-32 LoRA was merged into the base weights. The authors report 12,844 training actions across 12 manipulation benchmarks, 2 epochs, batch size 32 and a cosine learning-rate schedule starting at 5×10⁻⁵. CC BY 4.0 covers the annotations. Images remain subject to their source benchmarks’ licenses.
12844 actions · CC BY 4.0
- AgentRoboCritic-Bench · Evaluation · AgentRoboCritic-Dataset · Fine-tuning: AgentRoboCritic-Muse-Glimmer-30B
- AgentRoboCritic-Bench · Evaluation: AgentRoboCritic-Bench
- AgentRoboCritic-Dataset · Fine-tuning: AgentRoboCritic-Dataset
Full parameters and verification notes (model card)
The Hugging Face publisher is Chenyan Xiong Research Group at CMU. Project author affiliations are Carnegie Mellon University, University of Illinois at Urbana-Champaign, and Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI).
Supervised rank-32 LoRA was merged into the base weights. The authors report 12,844 training actions across 12 manipulation benchmarks, 2 epochs, batch size 32 and a cosine learning-rate schedule starting at 5×10⁻⁵.
The training and validation splits contain 12,844 and 1,548 actions, respectively. On the held-out evaluation the authors also report outcome Cohen’s κ of 0.444, failure-mode Cohen’s κ of 0.484, and an evaluation cost of USD 0.79.
Evaluation used temperature 0 and a maximum of 1,024 output tokens.