In this tutorial you will go through the full Human-in-the-Loop Sample-Efficient Reinforcement Learning (HIL-SERL) workflow using LeRobot. You will master training a policy with RL on a real robot in just a few hours.
HIL-SERL is a sample-efficient reinforcement learning algorithm that combines human demonstrations with online learning and human interventions. The approach starts from a small set of human demonstrations, uses them to train a reward classifier, and then employs an actor-learner architecture where humans can intervene during policy execution to guide exploration and correct unsafe behaviors. In this tutorial, you'll use a gamepad to provide interventions and control the robot during the learning process.
1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/scripts/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
HIL-SERL uses a modular processor pipeline architecture that processes robot observations and actions through a series of composable steps. The pipeline is divided into two main components:
#### Environment Processor Pipeline
The environment processor (`env_processor`) handles incoming observations and environment state:
1. **VanillaObservationProcessor**: Converts raw robot observations into standardized format
2. **JointVelocityProcessor** (optional): Adds joint velocity information to observations
3. **MotorCurrentProcessor** (optional): Adds motor current readings to observations
4. **ForwardKinematicsJointsToEE** (optional): Computes end-effector pose from joint positions
5. **ImageCropResizeProcessor** (optional): Crops and resizes camera images
6. **TimeLimitProcessor** (optional): Enforces episode time limits
7. **GripperPenaltyProcessor** (optional): Applies penalties for inappropriate gripper usage
8. **RewardClassifierProcessor** (optional): Automated reward detection using vision models
9. **ToBatchProcessor**: Converts data to batch format for neural network processing
10. **DeviceProcessor**: Moves data to the specified compute device (CPU/GPU)
#### Action Processor Pipeline
The action processor (`action_processor`) handles outgoing actions and human interventions:
1. **AddTeleopActionAsComplimentaryData**: Captures teleoperator actions for logging
2. **AddTeleopEventsAsInfo**: Records intervention events and episode control signals
3. **AddRobotObservationAsComplimentaryData**: Stores raw robot state for processing
4. **InterventionActionProcessor**: Handles human interventions and episode termination
- **InverseKinematicsEEToJoints**: Converts end-effector actions to joint targets
- **GripperVelocityToJoint**: Handles gripper control commands
#### Configuration Examples
**Basic Observation Processing**:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true,
"add_current_to_observation": false,
"display_cameras": false
}
}
}
}
```
**Image Processing**:
```json
{
"env": {
"processor": {
"image_preprocessing": {
"crop_params_dict": {
"observation.images.front": [180, 250, 120, 150],
"observation.images.side": [180, 207, 180, 200]
},
"resize_size": [128, 128]
}
}
}
}
```
**Inverse Kinematics Setup**:
```json
{
"env": {
"processor": {
"inverse_kinematics": {
"urdf_path": "path/to/robot.urdf",
"target_frame_name": "end_effector",
"end_effector_bounds": {
"min": [0.16, -0.08, 0.03],
"max": [0.24, 0.2, 0.1]
},
"end_effector_step_sizes": {
"x": 0.02,
"y": 0.02,
"z": 0.02
}
}
}
}
}
```
### Advanced Observation Processing
The HIL-SERL framework supports additional observation processing features that can improve policy learning:
#### Joint Velocity Processing
Enable joint velocity estimation to provide the policy with motion information:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true
}
}
}
}
```
This processor:
- Estimates joint velocities using finite differences between consecutive joint position readings
- Adds velocity information to the observation state vector
- Useful for policies that need motion awareness for dynamic tasks
#### Motor Current Processing
Monitor motor currents to detect contact forces and load conditions:
```json
{
"env": {
"processor": {
"observation": {
"add_current_to_observation": true
}
}
}
}
```
This processor:
- Reads motor current values from the robot's control system
- Adds current measurements to the observation state vector
- Helps detect contact events, object weights, and mechanical resistance
- Useful for contact-rich manipulation tasks
#### Combined Observation Processing
You can enable multiple observation processing features simultaneously:
```json
{
"env": {
"processor": {
"observation": {
"add_joint_velocity_to_observation": true,
"add_current_to_observation": true,
"add_ee_pose_to_observation": false,
"display_cameras": false
}
}
}
}
```
**Note**: Enabling additional observation features increases the state space dimensionality, which may require adjusting your policy network architecture and potentially collecting more training data.
Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
This helps simplify the problem of learning on the real robot in two ways: 1) by limiting the robot's operational space to a specific region that solves the task and avoids unnecessary or unsafe exploration, and 2) by allowing training in end-effector space rather than joint space. Empirically, learning in joint space for reinforcement learning in manipulation is often a harder problem - some tasks are nearly impossible to learn in joint space but become learnable when the action space is transformed to end-effector coordinates.
**Using find_joint_limits.py**
This script helps you find the safe operational bounds for your robot's end-effector. Given that you have a follower and leader arm, you can use the script to find the bounds for the follower arm that will be applied during training.
Bounding the action space will reduce the redundant exploration of the agent and guarantees safety.
```bash
python -m lerobot.scripts.find_joint_limits \
--robot.type=so100_follower \
--robot.port=/dev/tty.usbmodem58760431541 \
--robot.id=black \
--teleop.type=so100_leader \
--teleop.port=/dev/tty.usbmodem58760431551 \
--teleop.id=blue
```
**Workflow**
1. Run the script and move the robot through the space that solves the task
2. The script will record the minimum and maximum end-effector positions and the joint angles and prints them to the console, for example:
```
Max ee position [0.2417 0.2012 0.1027]
Min ee position [0.1663 -0.0823 0.0336]
Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
```
3. Use these values in the configuration of your teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field
**Example Configuration**
```json
"end_effector_bounds": {
"max": [0.24, 0.20, 0.10],
"min": [0.16, -0.08, 0.03]
}
```
### Collecting Demonstrations
With the bounds defined, you can safely collect demonstrations for training. Training RL with off-policy algorithm allows us to use offline datasets collected in order to improve the efficiency of the learning process.
**Setting Up Record Mode**
Create a configuration file for recording demonstrations (or edit an existing one like [env_config_so100.json](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json)):
Along with your robot, you will need a teleoperation device to control it in order to collect datasets of your task and perform interventions during the online training.
We support using a gamepad or a keyboard or the leader arm of the robot.
HIL-Serl learns actions in the end-effector space of the robot. Therefore, the teleoperation will control the end-effector's x,y,z displacements.
For that we need to define a version of the robot that takes actions in the end-effector space. Check the robot class `SO100FollowerEndEffector` and its configuration `SO100FollowerEndEffectorConfig` for the default parameters related to the end-effector space.
The SO101 leader arm has reduced gears that allows it to move and track the follower arm during exploration. Therefore, taking over is much smoother than the gearless SO100.
To setup the SO101 leader, you need to set the `control_mode` to `"leader"` and define the `teleop` section in the configuration file.
<p align="center"><i>SO101 leader teleoperation example, the leader tracks the follower, press `space` to intervene</i></p>
</details>
**Recording Demonstrations**
Start the recording process, an example of the config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json):
3. The episode ends with a reward of 1 when you press the "success" button
4. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
5. You can rerecord an episode by pressing the "rerecord" button
6. The process automatically continues to the next episode
7. After recording all episodes, the dataset is pushed to the Hugging Face Hub (optional) and saved locally
### Processing the Dataset
After collecting demonstrations, process them to determine optimal camera crops.
Reinforcement learning is sensitive to background distractions, so it is important to crop the images to the relevant workspace area.
Visual RL algorithms learn directly from pixel inputs, making them vulnerable to irrelevant visual information. Background elements like changing lighting, shadows, people moving, or objects outside the workspace can confuse the learning process. Good ROI selection should:
Most vision-based policies have been validated on square inputs of either **128×128** (default) or **64×64** pixels. We therefore advise setting the resize_size parameter to [128, 128] – or [64, 64] if you need to save GPU memory and bandwidth. Other resolutions are possible but have not been extensively tested.
The reward classifier plays an important role in the HIL-SERL workflow by automating reward assignment and automatically detecting episode success. Instead of manually defining reward functions or relying on human feedback for every timestep, the reward classifier learns to predict success/failure from visual observations. This enables the RL algorithm to learn efficiently by providing consistent and automated reward signals based on the robot's camera inputs.
This guide explains how to train a reward classifier for human-in-the-loop reinforcement learning implementation of LeRobot. Reward classifiers learn to predict the reward value given a state which can be used in an RL setup to train a policy.
**Note**: Training a reward classifier is optional. You can start the first round of RL experiments by annotating the success manually with your gamepad or keyboard device.
The reward classifier implementation in `modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
**Collecting a Dataset for the reward classifier**
Before training, you need to collect a dataset with labeled examples. The `record_dataset` function in `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
- **env.processor.reset.terminate_on_success**: Whether to automatically terminate episodes when success is detected (default: `true`)
- **env.fps**: Number of frames per second to record
- **dataset.push_to_hub**: Whether to push the dataset to the hub
The `env.processor.reset.terminate_on_success` parameter allows you to control episode termination behavior. When set to `false`, episodes will continue even after success is detected, allowing you to collect more positive examples with the reward=1 label. This is crucial for training reward classifiers as it provides more success state examples in your dataset. When set to `true` (default), episodes terminate immediately upon success detection.
**Important**: For reward classifier training, set `terminate_on_success: false` to collect sufficient positive examples. For regular HIL-SERL training, keep it as `true` to enable automatic episode termination when the task is completed successfully.
The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:
- **model_name**: Base model architecture (e.g., we mainly use `"helper2424/resnet10"`)
- **model_type**: `"cnn"` or `"transformer"`
- **num_cameras**: Number of camera inputs
- **num_classes**: Number of output classes (typically 2 for binary success/failure)
- **hidden_dim**: Size of hidden representation
- **dropout_rate**: Regularization parameter
- **learning_rate**: Learning rate for optimizer
Example configuration for training the [reward classifier](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/reward_classifier_train_config.json):
```json
{
"policy": {
"type": "reward_classifier",
"model_name": "helper2424/resnet10",
"model_type": "cnn",
"num_cameras": 2,
"num_classes": 2,
"hidden_dim": 256,
"dropout_rate": 0.1,
"learning_rate": 1e-4,
"device": "cuda",
"use_amp": true,
"input_features": {
"observation.images.front": {
"type": "VISUAL",
"shape": [3, 128, 128]
},
"observation.images.side": {
"type": "VISUAL",
"shape": [3, 128, 128]
}
}
}
}
```
**Training the Classifier**
To train the classifier, use the `train.py` script with your configuration:
The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
**Example Workflow for training the reward classifier**
1. **Create the configuration files**:
Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/tree/main).
The LeRobot system uses a distributed actor-learner architecture for training. This architecture decouples robot interactions from the learning process, allowing them to run concurrently without blocking each other. The actor server handles robot observations and actions, sending interaction data to the learner server. The learner server performs gradient descent and periodically updates the actor's policy weights. You will need to start two processes: a learner and an actor.
**Configuration Setup**
Create a training configuration file (example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_config_hilserl_so100.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
1. Configure the policy settings (`type="sac"`, `device`, etc.)
2. Set `dataset` to your cropped dataset
3. Configure environment settings with crop parameters
4. Check the other parameters related to SAC in [configuration_sac.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/sac/configuration_sac.py#L79).
- Execute rollouts of the policy to collect experience
- Sends transitions to the learner
- Receives updated policy parameters
**Training Flow**
The training proceeds automatically:
1. The actor executes the policy in the environment
2. Transitions are collected and sent to the learner
3. The learner updates the policy based on these transitions
4. Updated policy parameters are sent back to the actor
5. The process continues until the specified step limit is reached
**Human in the Loop**
- The key to learning efficiently is to have human interventions to provide corrective feedback and completing the task to aide the policy learning and exploration.
- To perform human interventions, you can press the upper right trigger button on the gamepad (or the `space` key on the keyboard). This will pause the policy actions and allow you to take over.
- A successful experiment is one where the human has to intervene at the start but then reduces the amount of interventions as the policy improves. You can monitor the intervention rate in the `wandb` dashboard.
- The figure shows the plot of the episodic reward over interaction step. The figure shows the effect of human interventions on the policy learning.
- The orange curve is an experiment without any human interventions. While the pink and blue curves are experiments with human interventions.
- We can observe that the number of steps where the policy starts achieving the maximum reward is cut by a quarter when human interventions are present.
**Monitoring and Debugging**
If you have `wandb.enable` set to `true` in your configuration, you can monitor training progress in real-time through the [Weights & Biases](https://wandb.ai/site/) dashboard.
The learning process is very sensitive to the intervention strategy. It will takes a few runs to understand how to intervene effectively. Some tips and hints:
- Allow the policy to explore for a few episodes at the start of training.
- Avoid intervening for long periods of time. Try to intervene in situation to correct the robot's behaviour when it goes off track.
- Once the policy starts achieving the task, even if its not perfect, you can limit your interventions to simple quick actions like a simple grasping commands.
The ideal behaviour is that your intervention rate should drop gradually during training as shown in the figure below.
Some configuration values have a disproportionate impact on training stability and speed:
- **`temperature_init`** (`policy.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
- **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
- **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.
