roboreason 0.1.4.8

Roboreason package

RoboReason

RoboReason is a python package that makes it easy to apply any reward model or video-language reasoning model to your robot videos.

Supported Models

• Robometer (https://robometer.github.io)
• SOLE-R1 (https://philipmit.github.io/sole-r1/)
• TOPReward (https://topreward.github.io/webpage/)
• RoboReward (https://arxiv.org/abs/2601.00675)
• OpenAI models (e.g., "GPT-5")
• Google models (e.g., "Gemini-3-Pro-Preview")

ToDos

• Enable fine-tuning of reward models on custom datasets

📦 File Structure

roboreason/
├── roboreason/         # Main package
│   ├── robometer/         # Robometer code
│   ├── sole.py            # SOLE-R1 code
│   ├── roboreward.py      # RoboReward code
│   ├── topreward.py       # TOPReward code
│   └── api_models.py      # OpenAI and Gemini APIs
├── test_videos/        # Example videos to test
├── model_outputs/      # Videos showing model outputs
├── docs/   
│   ├── lerobot_dataset_reward_annotation.mdx  # Examples showing integration with lerobot datasets
└── pyproject.toml      # Dependencies (uv)

Install

Option 1: quick pip install

pip install -U roboreason

Option 2: use uv for dependency management

1. Clone the repository:

git clone https://github.com/philipmit/roboreason

2. Install uv

pip install uv

3. Sync environment

uv sync

4. Activate environment

source .venv/bin/activate

---

Pre-download model checkpoints (optional)

# SOLE-R1 (8B)
python -c "from roboreason.utils.model_utils import get_model_dir; get_model_dir('sole-r1')"

# Robometer (4B)
python -c "from roboreason.utils.model_utils import get_model_dir; get_model_dir('robometer')"

# TOPReward (based on Qwen3-VL-8B)
python -c "from roboreason.utils.model_utils import get_model_dir; get_model_dir('topreward')"

# RoboReward (8B)
python -c "from roboreason.utils.model_utils import get_model_dir; get_model_dir('roboreward')"

Download all test videos from google drive link here (optional)

# pip install gdown
# cd /path/to/roboreason/test_videos/
gdown https://drive.google.com/drive/folders/1pXmiN-l8-khC4WABoMAn6saDvHjHGub0?usp=sharing

Download all videos showing example model outputs from google drive link here (optional)

# pip install gdown
# cd /path/to/roboreason/model_outputs/
gdown https://drive.google.com/drive/folders/1gi-sTk8JssO9_UO6dHTqnkeAMjyW0YZb?usp=sharing

---

Quick start: Example reward generation and plotting

# pip install -U roboreason
import roboreason as rr

video_paths = ['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4']
task_description="Pick up the cube from the table."

# Robometer
rewards, success_probs = rr.generate(model="Robometer",  task_description=task_description, video_paths=video_paths, view_type_per_video=['external'])
output_robometer = {"model": "Robometer", "rewards": rewards[0]}

# SOLE-R1
rewards, reasoning_traces = rr.generate(model="SOLE-R1",  task_description=task_description, video_paths=video_paths, view_type_per_video=['external and wrist'])
output_sole = {"model": "SOLE-R1", "rewards": rewards[0], "reasoning_traces": reasoning_traces[0]}

# Ground-truth rewards (available for test videos from sim environments)
import json
with open(video_paths[0].replace(".mp4", "/data.json"), 'r') as f:
    data = json.load(f)

output_groundtruth = {"model": "groundtruth", "rewards": data['ground-truth rewards']}

rr.video_plot(outputs=[output_groundtruth, output_sole, output_robometer], plot_save_path='model_outputs/combined/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4', video_path = video_paths[0])

---

Examples for generating across all models

Robometer

import roboreason as rr

rewards, success_probs = rr.generate(
    model="Robometer",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external'],
    verbose=False
)

SOLE-R1

import roboreason as rr

rewards, reasoning_traces = rr.generate(
    model="SOLE-R1",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external and wrist'],
    verbose=False
)

TOPReward

import roboreason as rr

rewards = rr.generate(
    model="TOPReward",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external'],
    verbose=False
)

RoboReward

import roboreason as rr

rewards = rr.generate(
    model="RoboReward",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external'],
    verbose=False
)

GPT-5 (and other OpenAI models)

import roboreason as rr

# requires OpenAI API key: https://developers.openai.com/api/docs/quickstart
API_KEY = "..."

rewards, reasoning_traces = rr.generate(
    model="GPT-5",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external'], 
    key=API_KEY, 
    verbose=False
)

Gemini-3-Pro (and other Google models)

import roboreason as rr

# requires Gemini API key: https://ai.google.dev/gemini-api/docs/api-key
API_KEY = "..."

rewards, reasoning_traces = rr.generate(
    model="Gemini-3-Pro-Preview",  
    task_description="Pick up the cube from the table.", 
    video_paths=['test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4'], 
    view_type_per_video=['external'], 
    key=API_KEY,
    verbose=False
)

Video plotting

import roboreason as rr

# Robometer
rewards, success_probs = rr.generate(model="Robometer",  task_description=task_description, video_paths=video_paths, view_type_per_video=['external'])
output_robometer = {"model": "Robometer", "rewards": rewards[0]}

# SOLE-R1
rewards, reasoning_traces = rr.generate(model="SOLE-R1",  task_description=task_description, video_paths=video_paths, view_type_per_video=['external and wrist'])
output_sole = {"model": "SOLE-R1", "rewards": rewards[0], "reasoning_traces": reasoning_traces[0]}

# Ground-truth rewards (available for test videos from sim environments)
import json
with open(video_paths[0].replace(".mp4", "/data.json"), 'r') as f:
    data = json.load(f)

output_groundtruth = {"model": "groundtruth", "rewards": data['ground-truth rewards']}

rr.video_plot(
    outputs=[output_sole, output_robometer], 
    plot_save_path='model_outputs/combined/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4', 
    video_path = 'test_videos/robosuite/lift/unsuccessful/robosuite_lift_episode_11_unsuccessful_max_reward_37.mp4',
    verbose=False
)

---

rr.generate

Argument	Type	Required	Description
model	str	✅	Name of the model to use. Options include: "Robometer", "SOLE-R1", "TOPReward", "RoboReward", OpenAI models (e.g."GPT-5"), Google models (e.g., "Gemini-3-Pro-Preview")
task_description	str	✅	Natural language description of the task the robot is performing.
video_paths	List[str]	✅	List of paths to input video files.
view_type_per_video	List[str]	✅	List specifying the camera view(s) used for reward reasoning for each video (e.g., "external", "wrist", or "external and wrist").
key	str	❌	API key required for external models (e.g., OpenAI or Gemini). Not needed for local models.

Model Type	Return Values
SOLE-R1 / GPT / Gemini	rewards, reasoning_traces
Robometer	rewards, success_probs
TOPReward / RoboReward	rewards

rr.video_plot

Argument	Type	Required	Description
outputs	List[dict]	❌*	List of model outputs (e.g., from rr.generate) to visualize together.
plot_save_path	str	❌	Path where the output video with overlays will be saved.
video_path	str	❌	Path to the original video file being visualized.
view_type	str	❌	View type used for visualization (e.g., "external", "wrist", "external and wrist").
show_reasoning_traces	bool	❌	Whether to overlay reasoning traces on the video. Default: False.
show_all_frames	bool	❌	Whether to render all frames instead of sampled frames. Default: False.
model	str	❌**	Model name (used when calling video_plot directly instead of passing outputs).
task_description	str	❌**	Task description (used in direct-call mode).
video_paths	List[str]	❌**	Input videos (used in direct-call mode).
view_type_per_video	List[str]	❌**	View types per video (used in direct-call mode).
key	str	❌**	API key (if required for model).

---

Acknowledgements

RoboReason builds upon the following repos:

• RewardScope (https://github.com/philfung/reward-scope)
• Robometer (https://github.com/robometer/robometer)
• TOPReward (https://github.com/TOPReward/TOPReward)