AM-ACT#

本节介绍 LeRobot 的通用功能。请按目标机器人、数据集和运行环境配置示例；用于 AlohaMini 2 / 2 Pro 前，需要确认硬件与接口兼容性。

AM-ACT extends ACT with optional per-dimension classification heads while keeping all other action dimensions as continuous L1 regression outputs. It also supports fixed action dimensions, optional loss groups, visual-only inputs, selecting a subset of observation.state, and post-unnormalization action scaling.

Minimal training command#

conda activate lerobot_alohamini

python -m lerobot.scripts.lerobot_train \
  --dataset.repo_id=local/my_dataset \
  --dataset.root=/path/to/my_dataset \
  --policy.type=am_act \
  --policy.device=cuda \
  --policy.fixed_action_dims='[]' \
  --policy.discrete_action_dims='[14,15,16]' \
  --policy.discrete_action_values='[[-0.15,0,0.15],[-0.15,0,0.15],[-45,0,45]]' \
  --policy.discrete_action_class_weights='[[3,1,1.5],[3,1,2],[2,1,2]]' \
  --policy.discrete_action_loss_weight=1.0 \
  --output_dir=outputs/am_act_run \
  --steps=100000 \
  --batch_size=2

The values above are only an example. Dimension indices, physical class values, and class weights must match the dataset.

Main parameters#

Parameter	Meaning
`fixed_action_dims`	Action dimensions excluded from training and forced to zero in normalized space.
`discrete_action_dims`	Action dimensions trained with classification instead of L1 regression. Empty by default.
`discrete_action_values`	Physical class values for each discrete dimension. Their order defines the class order.
`discrete_action_class_weights`	Optional cross-entropy weights in the same order as `discrete_action_values`.
`discrete_action_loss_weight`	Multiplier applied to the mean classification loss.
`action_loss_groups`	Optional named groups used to calculate separately normalized continuous L1 losses.
`action_loss_weights`	Optional weights for the named continuous loss groups.
`observation_state_dims`	Optional subset of `observation.state`; omit `observation.state` from the dataset features for visual-only training.
`inference_action_scale_dims`	Physical action dimensions scaled after unnormalization during inference.
`inference_action_scale`	Scale applied to `inference_action_scale_dims`.

discrete_action_values are specified in physical dataset units. AM-ACT converts them to normalized class centers using the active dataset statistics and decodes predicted classes before the standard action unnormalizer.

Load a trained checkpoint#

Use the saved policy like any other LeRobot policy:

conda activate lerobot_alohamini

python -m lerobot.scripts.lerobot_eval \
  --policy.path=outputs/am_act_run/checkpoints/last/pretrained_model \
  --env.type=<environment_type>

For robot-specific evaluation scripts, pass the same checkpoint path through their --policy.path option.