Policies and training

This page summarizes policies and models for learning and development. For the initial training of AlohaMini, you can still verify the data and equipment along the ACT Main Line; other policies are configured according to their respective dependencies, input formats, and running interfaces.

1. Select the reading entrance

policy/Topic	Tutorial entrance	Reading highlights
ACT	Complete tutorial	action chunk, training and evaluation
AM-ACT	Custom implementation instructions	Discrete action dimensions, fixed dimensions, loss groups, and output scaling
Diffusion	Implementation instructions	Implementation and parameter description
SmolVLA	Complete tutorial	Model dependencies, fine-tuning, inference and data
Pi0 / Pi0.5	Pi0、Pi0.5	LeRobot’s built-in implementation is different from Standalone OpenPI
Pi0 FAST	Complete tutorial	Action representation and model use
GR00T	Complete tutorial	Modality, configuration and training
RTC	Complete tutorial	Asynchronous action blocks and compatibility policies
Multitasking DiT	Complete tutorial	Multitasking configuration and data
FastWAM	Complete tutorial	Models and exclusive dependencies
MolmoAct2	Complete tutorial	Model input, fine-tuning and use
VLA-JEPA	Complete tutorial	Configure, train and deploy
EO-1 / Evo-1	EO-1、Evo-1	Corresponding implementation version and dependencies
LingBot-VA / Wall-OSS / XVLA	LingBot-VA、Wall-OSS、XVLA	Data and interfaces specific to each model
TD-MPC / VQ-BeT / SARM	TD-MPC、VQ-BeT、SARM	Part of the README is for reference, please read the corresponding text and code carefully.

View all policy tutorials. Before using other policies, please confirm the input, action dimensions and inference interface of AlohaMini, and complete the real machine verification.

2. Before moving from ACT data to other policies

1. Keep a copy of the validated dataset and ACT baseline documenting cameras, feature order, action units, and calibration.

2. Install the corresponding extras according to the model tutorial, and check the basic weights and usage conditions.

3. Confirm that the image name, language task field, state/action dimension, and normalized statistics are consistent with the model requirements.

4. Choose independent output directories and experiment names to avoid overwriting existing checkpoints.

5. First run a small-scale data loading and training check to confirm that the log, loss, and save results are normal.

6. Check the inference interface, control frequency, and action block policy when deploying; ACT does not use the RTC path.

The example batch size, number of training steps, and memory records depend on the specific model and training tasks and cannot be used as resource commitments for all devices.

3. Unique parameters of AM-ACT

AM-ACT allows some action dimensions to be categorical, while the remaining dimensions are still trained as continuous values.

parameters	function
fixed_action_dims	Dimensions that are excluded from training and fixed to zero in the normalized space
discrete_action_dims	Dimensional indexing using categorical predictions
discrete_action_values	The physical action values corresponding to each category are in the same order as the category.
discrete_action_class_weights	Corresponding category weight
discrete_action_loss_weight	Classification loss ratio
action_loss_groups / action_loss_weights	Continuous dimension grouping and group weights
observation_state_dims	Select a subset of input states
inference_action_scale_dims / inference_action_scale	Scale the specified output after denormalization

The example configuration uses [14,15,16] as a discrete dimension example; whether it corresponds to the chassis should be judged based on the actual data feature sequence. Zero in normalized space also does not necessarily equal physical zero action. For complete training commands and loading methods, see AM-ACT detailed configuration.

4. Data and training tools

• Dataset v3 format, Migrate old data

• Dataset editing tools, action expression

• PEFT training, Multi-GPU training

• inference Tutorial, Asynchronous inference

• Human-in-the-loop data collection, HIL-SERL

[Tutorial Library] Summarizes various topics (tutorial-library.md), including other robot and simulation benchmark examples.