LiberalMind Panda Policy
Learning to control a robot through artificial intelligence.
LiberalMind is an AI research project exploring autonomous robotic control and embodied intelligence.
This repository contains the trained best.pt checkpoint for our Panda robot policy, intended for experimentation and evaluation in robotic simulation environments.
Our goal is to bridge the gap between neural network intelligence and physical interaction with the world.
Project Overview
| Component | Details |
|---|---|
| Organization | LiberalMind |
| Robot | Franka Emika Panda |
| Model format | PyTorch .pt |
| Checkpoint | best.pt |
| Simulation | PyBullet |
| Research area | Embodied AI / Robotics |
| Project status | Research |
The Robot
The target platform is the Franka Emika Panda, a 7-DoF robotic manipulator with a parallel gripper.
The simulation provides an environment for testing robot movements, control policies, and autonomous manipulation.
Repository Structure
panda-policy/
β
βββ README.md
βββ best.pt
Quick Start
1. Install dependencies
We recommend Python 3.10 or newer.
python -m venv .venv
source .venv/bin/activate
pip install torch huggingface_hub pybullet
2. Download the checkpoint
For private repositories, first authenticate with Hugging Face:
hf auth login
Then download the trained weights:
from huggingface_hub import hf_hub_download
checkpoint_path = hf_hub_download(
repo_id="LiberalMind/panda-policy",
filename="best.pt",
token=True
)
print("Checkpoint downloaded:", checkpoint_path)
3. Load the model checkpoint
import torch
checkpoint = torch.load(
checkpoint_path,
map_location="cpu",
weights_only=False
)
print("Checkpoint loaded successfully.")
if isinstance(checkpoint, dict):
print("Available keys:", list(checkpoint.keys()))
Note: The checkpoint is a set of learned weights. To execute the trained policy, the original model architecture, observation preprocessing, and action mapping from the training code must also be available.
4. Launch the Panda simulation
The following script launches a Franka Panda robot in PyBullet.
import pybullet as p
import pybullet_data
import time
# Initialize physics
physics_client = p.connect(p.GUI)
p.setAdditionalSearchPath(
pybullet_data.getDataPath()
)
p.setGravity(0, 0, -9.81)
p.setTimeStep(1 / 240)
# Ground
p.loadURDF("plane.urdf")
# Franka Panda
robot = p.loadURDF(
"franka_panda/panda.urdf",
[0, 0, 0],
useFixedBase=True
)
# Camera
p.resetDebugVisualizerCamera(
cameraDistance=1.5,
cameraYaw=45,
cameraPitch=-30,
cameraTargetPosition=[0, 0, 0.5]
)
# Default arm pose
joint_indices = list(range(7))
home_positions = [
0.0,
-0.215,
0.0,
-2.57,
0.0,
2.356,
2.753
]
p.setJointMotorControlArray(
robot,
joint_indices,
p.POSITION_CONTROL,
targetPositions=home_positions,
forces=[87] * 7
)
print("LiberalMind Panda simulation started.")
while p.isConnected():
p.stepSimulation()
time.sleep(1 / 240)
This is a simulation smoke test using a predefined joint pose, not a trained-policy rollout.
Policy Integration
A complete autonomous rollout requires the following pipeline:
Simulation Environment
|
v
Observation
|
v
Neural Network Policy
|
v
Action Prediction
|
v
Robot Joint Control
|
v
Physics Simulation
|
v
Next State
|
+----------+
|
Repeat
The policy must use the exact observation format, normalization, output dimensions, and control conventions used during training.
Research Roadmap
- Train and save a model checkpoint
- Publish model weights on Hugging Face
- Reconstruct the policy architecture
- Connect the policy to the Panda simulator
- Evaluate autonomous manipulation
- Test generalization across environments
- Explore transfer from simulation to real robotic hardware
About LiberalMind
LiberalMind is an independent AI research initiative focused on developing intelligent models and exploring the next generation of autonomous systems.
We are interested in the intersection of:
- Artificial Intelligence
- Embodied Intelligence
- Robotics
- Autonomous Agents
- Reinforcement Learning
- Real-World AI Applications
Our long-term vision is to develop AI systems capable of reasoning, learning, and interacting with the physical world.
Organization: LiberalMind
Model: LiberalMind/panda-policy
Framework: PyTorch
Repository: Hugging Face
Building intelligence beyond the screen.