Skip to content
Computer vision

AgentRoboCritic-Muse-Glimmer-30B

A critic for robot actions

2 min readupdated

The model evaluates whether a robot-performed action fulfilled a given local intent based on the task, agent command, and visual observations from multiple cameras. It outputs a structured JSON response containing a verdict, failure category, and explanation derived from what it saw. The model is built on meta-models/Muse-Glimmer-30B (size 29.78 billion parameters) using LoRA (a small add-on to the base model) followed by weight merging. It is useful for researchers and developers in robotics and autonomous agents for analyzing manipulation errors.

Key features

  • Analyzes visual observations and task context to evaluate robot actions.
  • Returns structured JSON with success fields and failure reasons.
  • Trained on 12,844 actions from 12 manipulation benchmarks annotated with GPT-6 Astra.
  • Shows accuracy comparable to large closed models on test sets.

What you can do

  • Apply in robotic agent evaluation pipelines for automated error detection.
  • Use for analyzing failed grasps and classifying manipulation failures.
  • Test the model's judgment quality on your own robotics episodes.

Example

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

checkpoint = "cx-cmu/AgentRoboCritic-Muse-Glimmer-30B"
processor = AutoProcessor.from_pretrained(checkpoint)
critic = AutoModelForImageTextToText.from_pretrained(
    checkpoint, dtype=torch.bfloat16, device_map="auto"
).eval()
Read moreDetails
Benchmarks (test on 400 actions): overall accuracy 67.75%, outcome accuracy 76.25%, failure mode accuracy 74.24%
Hardware requirements: a machine with sufficient memory to run a 30B model in bfloat16
Training: LoRA (rank 32), 2 epochs, batch 32, learning rate 5×10&sup5; with cosine schedule
License: Apache 2.0 (based on Muse-Glimmer-30B with its own usage policy)
Discuss Telegram 0 13

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 55 GB · ready files62 GB
  • FP8 / INT8recommendedweights 28 GB31 GB
  • 4-bitweights 16 GB18 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # cx-cmu/AgentRoboCritic-Muse-Glimmer-30B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download cx-cmu/AgentRoboCritic-Muse-Glimmer-30B --local-dir ~/models/AgentRoboCritic-Muse-Glimmer-30B
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Muse-Glimmer-30B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face
Merges 1All on Hugging Face

About the model

cx-cmu/AgentRoboCritic-Muse-Glimmer-30B
Author
cx-cmu
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

130

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments