hangX/lerobot
1
0
Fork 0
forked from tangger/lerobot
Code Pull Requests Activity

Removed depleted files and scripts

Browse Source
This commit is contained in:
Michel Aractingi
2025-03-24 14:41:01 +01:00
parent 1c8daf11fd
commit 2abbd60a0d
15 changed files with 0 additions and 10042 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
lerobot/configs/default.yaml
View File

@@ -1,131 +0,0 @@
defaults:
- _self_
- env: pusht
- policy: diffusion
- robot: so100
hydra:
run:
# Set `dir` to where you would like to save all of the run outputs. If you run another training session
# with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
dir: outputs/train/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
job:
name: default
# Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure
# `hydra.run.dir` is the directory of an existing run with at least one checkpoint in it.
# Note that when resuming a run, the default behavior is to use the configuration from the checkpoint,
# regardless of what's provided with the training command at the time of resumption.
resume: false
device: cuda # cpu
# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
# automatic gradient scaling is used.
use_amp: false
# `seed` is used for training (eg: model initialization, dataset shuffling)
# AND for the evaluation environments.
seed: ???
# You may provide a list of datasets here. `train.py` creates them all and concatenates them. Note: only data
# keys common between the datasets are kept. Each dataset gets and additional transform that inserts the
# "dataset_index" into the returned item. The index mapping is made according to the order in which the
# datsets are provided.
dataset_repo_id: lerobot/pusht
video_backend: pyav
training:
offline_steps: ???
# Number of workers for the offline training dataloader.
num_workers: 4
batch_size: ???
eval_freq: ???
log_freq: 200
save_checkpoint: true
# Checkpoint is saved every `save_freq` training iterations and after the last training step.
save_freq: ???
# Online training. Note that the online training loop adopts most of the options above apart from the
# dataloader options. Unless otherwise specified.
# The online training look looks something like:
#
# for i in range(online_steps):
# do_online_rollout_and_update_online_buffer()
# for j in range(online_steps_between_rollouts):
# batch = next(dataloader_with_offline_and_online_data)
# loss = policy(batch)
# loss.backward()
# optimizer.step()
#
online_steps: ???
# How many episodes to collect at once when we reach the online rollout part of the training loop.
online_rollout_n_episodes: 1
# The number of environments to use in the gym.vector.VectorEnv. This ends up also being the batch size for
# the policy. Ideally you should set this to by an even divisor or online_rollout_n_episodes.
online_rollout_batch_size: 1
# How many optimization steps (forward, backward, optimizer step) to do between running rollouts.
online_steps_between_rollouts: null
# The proportion of online samples (vs offline samples) to include in the online training batches.
online_sampling_ratio: 0.5
# First seed to use for the online rollout environment. Seeds for subsequent rollouts are incremented by 1.
online_env_seed: null
# Sets the maximum number of frames that are stored in the online buffer for online training. The buffer is
# FIFO.
online_buffer_capacity: null
# The minimum number of frames to have in the online buffer before commencing online training.
# If online_buffer_seed_size > online_rollout_n_episodes, the rollout will be run multiple times until the
# seed size condition is satisfied.
online_buffer_seed_size: 0
# Whether to run the online rollouts asynchronously. This means we can run the online training steps in
# parallel with the rollouts. This might be advised if your GPU has the bandwidth to handle training
# + eval + environment rendering simultaneously.
do_online_rollout_async: false
image_transforms:
# These transforms are all using standard torchvision.transforms.v2
# You can find out how these transformations affect images here:
# https://pytorch.org/vision/0.18/auto_examples/transforms/plot_transforms_illustrations.html
# We use a custom RandomSubsetApply container to sample them.
# For each transform, the following parameters are available:
# weight: This represents the multinomial probability (with no replacement)
# used for sampling the transform. If the sum of the weights is not 1,
# they will be normalized.
# min_max: Lower & upper bound respectively used for sampling the transform's parameter
# (following uniform distribution) when it's applied.
# Set this flag to `true` to enable transforms during training
enable: false
# This is the maximum number of transforms (sampled from these below) that will be applied to each frame.
# It's an integer in the interval [1, number of available transforms].
max_num_transforms: 3
# By default, transforms are applied in Torchvision's suggested order (shown below).
# Set this to True to apply them in a random order.
random_order: false
brightness:
weight: 1
min_max: [0.8, 1.2]
contrast:
weight: 1
min_max: [0.8, 1.2]
saturation:
weight: 1
min_max: [0.5, 1.5]
hue:
weight: 1
min_max: [-0.05, 0.05]
sharpness:
weight: 1
min_max: [0.8, 1.2]
eval:
n_episodes: 1
# `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
batch_size: 1
# `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
use_async_envs: false
wandb:
enable: false
# Set to true to disable saving an artifact despite save_checkpoint == True
disable_artifact: false
project: lerobot
notes: ""

30
lerobot/configs/env/maniskill_example.yaml vendored
View File

@@ -1,30 +0,0 @@
# @package _global_
fps: 400
env:
name: maniskill/pushcube
task: PushCube-v1
image_size: 64
control_mode: pd_ee_delta_pose
state_dim: 25
action_dim: 7
fps: ${fps}
obs: rgb
render_mode: rgb_array
render_size: 64
device: cuda
reward_classifier:
pretrained_path: null
config_path: null
wrapper:
joint_masking_action_space: null
delta_action: null
video_record:
enabled: false
record_dir: maniskill_videos
trajectory_name: trajectory
fps: ${fps}

50
lerobot/configs/env/so100_real.yaml vendored
View File

@@ -1,50 +0,0 @@
# @package _global_
fps: 10
env:
name: real_world
task: null
state_dim: 15
action_dim: 3
fps: ${fps}
device: mps
wrapper:
crop_params_dict:
observation.images.front: [171, 207, 116, 251]
observation.images.side: [232, 200, 142, 204]
resize_size: [128, 128]
control_time_s: 10
reset_follower_pos: false
use_relative_joint_positions: true
reset_time_s: 5
display_cameras: false
delta_action: null #0.3
joint_masking_action_space: null #[1, 1, 1, 1, 0, 0] # disable wrist and gripper
add_joint_velocity_to_observation: true
add_ee_pose_to_observation: true
# If null then the teleoperation will be used to reset the robot
# Bounds for pushcube_gamepad_lerobot15 dataset and experiments
# fixed_reset_joint_positions: [-19.86, 103.19, 117.33, 42.7, 13.89, 0.297]
# ee_action_space_params: # If null then ee_action_space is not used
# bounds:
# max: [0.291, 0.147, 0.074]
# min: [0.139, -0.143, 0.03]
# Bounds for insertcube_gamepad dataset and experiments
fixed_reset_joint_positions: [20.0, 90., 90., 75., -0.7910156, -0.5673759]
ee_action_space_params:
bounds:
max: [0.25295413, 0.07498981, 0.06862044]
min: [0.2010096, -0.12, 0.0433196]
use_gamepad: true
x_step_size: 0.03
y_step_size: 0.03
z_step_size: 0.03
reward_classifier:
pretrained_path: null # outputs/classifier/13-02-random-sample-resnet10-frozen/checkpoints/best/pretrained_model
config_path: null # lerobot/configs/policy/hilserl_classifier.yaml

61
lerobot/configs/policy/hilserl_classifier.yaml
View File

@@ -1,61 +0,0 @@
# @package _global_
defaults:
- _self_
hydra:
run:
# Set `dir` to where you would like to save all of the run outputs. If you run another training session
# with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
dir: outputs/train_hilserl_classifier/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${hydra.job.name}
job:
name: default
seed: 13
dataset_repo_id: aractingi/push_cube_square_light_reward_cropped_resized
# aractingi/push_cube_square_reward_1_cropped_resized
dataset_root: data/aractingi/push_cube_square_light_reward_cropped_resized
local_files_only: true
train_split_proportion: 0.8
# Required by logger
env:
name: "classifier"
task: "binary_classification"
training:
num_epochs: 6
batch_size: 16
learning_rate: 1e-4
num_workers: 4
grad_clip_norm: 10
use_amp: true
log_freq: 1
eval_freq: 1 # How often to run validation (in epochs)
save_freq: 1 # How often to save checkpoints (in epochs)
save_checkpoint: true
image_keys: ["observation.images.front", "observation.images.side"]
label_key: "next.reward"
profile_inference_time: false
profile_inference_time_iters: 20
eval:
batch_size: 16
num_samples_to_log: 30 # Number of validation samples to log in the table
policy:
name: "hilserl/classifier"
model_name: "helper2424/resnet10" # "facebook/convnext-base-224
model_type: "cnn"
num_cameras: 2 # Has to be len(training.image_keys)
wandb:
enable: false
project: "classifier-training"
job_name: "classifier_training_0"
disable_artifact: false
device: "mps"
resume: false
output_dir: "outputs/classifier/old_trainer_resnet10_frozen"

118
lerobot/configs/policy/sac_maniskill.yaml
View File

@@ -1,118 +0,0 @@
# @package _global_
# Train with:
#
# python lerobot/scripts/train.py \
# +dataset=lerobot/pusht_keypoints
# env=pusht \
# env.gym.obs_type=environment_state_agent_pos \
seed: 1
# dataset_repo_id: "AdilZtn/Maniskill-Pushcube-demonstration-medium"
dataset_repo_id: null
training:
# Offline training dataloader
num_workers: 4
batch_size: 512
grad_clip_norm: 40.0
lr: 3e-4
storage_device: "cuda"
eval_freq: 2500
log_freq: 10
save_freq: 1000000
online_steps: 1000000
online_rollout_n_episodes: 10
online_rollout_batch_size: 10
online_steps_between_rollouts: 1000
online_sampling_ratio: 1.0
online_env_seed: 10000
online_buffer_capacity: 200000
offline_buffer_capacity: 100000
online_buffer_seed_size: 0
online_step_before_learning: 500
do_online_rollout_async: false
policy_update_freq: 1
policy:
name: sac
pretrained_model_path:
# Input / output structure.
n_action_repeats: 1
horizon: 1
n_action_steps: 1
shared_encoder: true
# vision_encoder_name: "helper2424/resnet10"
vision_encoder_name: null
# freeze_vision_encoder: true
freeze_vision_encoder: false
input_shapes:
# # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
observation.state: ["${env.state_dim}"]
observation.image: [3, 64, 64]
output_shapes:
action: [7]
camera_number: 1
# Normalization / Unnormalization
# input_normalization_modes: null
input_normalization_modes:
observation.state: min_max
observation.image: mean_std
# input_normalization_params: null
input_normalization_params:
observation.state:
min: [-1.9361e+00, -7.7640e-01, -7.7094e-01, -2.9709e+00, -8.5656e-01,
1.0764e+00, -1.2680e+00, 0.0000e+00, 0.0000e+00, -9.3448e+00,
-3.3828e+00, -3.8420e+00, -5.2553e+00, -3.4154e+00, -6.5082e+00,
-6.0500e+00, -8.7193e+00, -8.2337e+00, -3.4650e-01, -4.9441e-01,
8.3516e-03, -3.1114e-01, -9.9700e-01, -2.3471e-01, -2.7137e-01]
max: [ 0.8644, 1.4306, 1.8520, -0.7578, 0.9508, 3.4901, 1.9381, 0.0400,
0.0400, 5.0885, 4.7156, 7.9393, 7.9100, 2.9796, 5.7720, 4.7163,
7.8145, 9.7415, 0.2422, 0.4505, 0.6306, 0.2622, 1.0000, 0.5135,
0.4001]
observation.image:
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
output_normalization_modes:
action: min_max
output_normalization_params:
action:
min: [-0.03, -0.03, -0.03, -0.03, -0.03, -0.03, -0.03]
max: [0.03, 0.03, 0.03, 0.03, 0.03, 0.03, 0.03]
output_normalization_shapes:
action: [7]
# Architecture / modeling.
# Neural networks.
image_encoder_hidden_dim: 32
# discount: 0.99
discount: 0.80
temperature_init: 1.0
num_critics: 2 #10
num_subsample_critics: null
critic_lr: 3e-4
actor_lr: 3e-4
temperature_lr: 3e-4
# critic_target_update_weight: 0.005
critic_target_update_weight: 0.01
utd_ratio: 2 # 10
actor_learner_config:
learner_host: "127.0.0.1"
learner_port: 50051
policy_parameters_push_frequency: 4
concurrency:
actor: 'threads'
learner: 'threads'

89
lerobot/configs/policy/sac_pusht_keypoints.yaml
View File

@@ -1,89 +0,0 @@
# @package _global_
# Train with:
#
# python lerobot/scripts/train.py \
# env=pusht \
# +dataset=lerobot/pusht_keypoints
seed: 1
dataset_repo_id: lerobot/pusht_keypoints
training:
offline_steps: 0
# Offline training dataloader
num_workers: 4
batch_size: 128
grad_clip_norm: 10.0
lr: 3e-4
eval_freq: 50000
log_freq: 500
save_freq: 50000
online_steps: 1000000
online_rollout_n_episodes: 10
online_rollout_batch_size: 10
online_steps_between_rollouts: 1000
online_sampling_ratio: 1.0
online_env_seed: 10000
online_buffer_capacity: 40000
online_buffer_seed_size: 0
do_online_rollout_async: false
delta_timestamps:
observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
action: "[i / ${fps} for i in range(${policy.horizon})]"
next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
policy:
name: sac
pretrained_model_path:
# Input / output structure.
n_action_repeats: 1
horizon: 5
n_action_steps: 5
input_shapes:
# TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
observation.environment_state: [16]
observation.state: ["${env.state_dim}"]
output_shapes:
action: ["${env.action_dim}"]
# Normalization / Unnormalization
input_normalization_modes:
observation.environment_state: min_max
observation.state: min_max
output_normalization_modes:
action: min_max
# Architecture / modeling.
# Neural networks.
# image_encoder_hidden_dim: 32
discount: 0.99
temperature_init: 1.0
num_critics: 2
num_subsample_critics: None
critic_lr: 3e-4
actor_lr: 3e-4
temperature_lr: 3e-4
critic_target_update_weight: 0.005
utd_ratio: 2
# # Loss coefficients.
# reward_coeff: 0.5
# expectile_weight: 0.9
# value_coeff: 0.1
# consistency_coeff: 20.0
# advantage_scaling: 3.0
# pi_coeff: 0.5
# temporal_decay_coeff: 0.5
# # Target model.
# target_model_momentum: 0.995

120
lerobot/configs/policy/sac_real.yaml
View File

@@ -1,120 +0,0 @@
# @package _global_
# Train with:
#
# python lerobot/scripts/train.py \
# +dataset=lerobot/pusht_keypoints
# env=pusht \
# env.gym.obs_type=environment_state_agent_pos \
seed: 1
dataset_repo_id: aractingi/insertcube_simple
training:
# Offline training dataloader
num_workers: 4
# batch_size: 256
batch_size: 512
grad_clip_norm: 10.0
lr: 3e-4
eval_freq: 2500
log_freq: 1
save_freq: 2000000
online_steps: 1000000
online_rollout_n_episodes: 10
online_rollout_batch_size: 10
online_steps_between_rollouts: 1000
online_sampling_ratio: 1.0
online_env_seed: 10000
online_buffer_capacity: 10000
online_buffer_seed_size: 0
online_step_before_learning: 100 #5000
do_online_rollout_async: false
policy_update_freq: 1
# delta_timestamps:
# observation.environment_state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
# observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
# action: "[i / ${fps} for i in range(${policy.horizon})]"
# next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
policy:
name: sac
pretrained_model_path:
# Input / output structure.
n_action_repeats: 1
horizon: 1
n_action_steps: 1
shared_encoder: true
vision_encoder_name: "helper2424/resnet10"
freeze_vision_encoder: true
input_shapes:
# # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
observation.state: ["${env.state_dim}"]
observation.images.front: [3, 128, 128]
observation.images.side: [3, 128, 128]
# observation.image: [3, 128, 128]
output_shapes:
action: ["${env.action_dim}"]
# Normalization / Unnormalization
input_normalization_modes:
observation.images.front: mean_std
observation.images.side: mean_std
observation.state: min_max
input_normalization_params:
observation.images.front:
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
observation.images.side:
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
observation.state:
# 6- joint positions, 6- joint velocities, 3- ee position
max: [ 52.822266, 136.14258, 142.03125, 72.1582, 22.675781, -0.5673759, 100., 100., 100., 100., 100., 100., 0.25295413, 0.07498981, 0.06862044]
min: [-2.6367188, 86.572266, 89.82422, 12.392578, -26.015625, -0.5673759, -100., -100., -100., -100., -100., -100., 0.2010096, -0.12, 0.0433196]
output_normalization_modes:
action: min_max
output_normalization_params:
action:
min: [-0.03, -0.03, -0.01]
max: [0.03, 0.03, 0.03]
# Architecture / modeling.
# Neural networks.
image_encoder_hidden_dim: 32
# discount: 0.99
discount: 0.97
temperature_init: 1.0
num_critics: 2 #10
camera_number: 2
num_subsample_critics: null
critic_lr: 3e-4
actor_lr: 3e-4
temperature_lr: 3e-4
# critic_target_update_weight: 0.005
critic_target_update_weight: 0.01
utd_ratio: 2 # 10
actor_learner_config:
learner_host: "127.0.0.1"
learner_port: 50051
policy_parameters_push_frequency: 15
# # Loss coefficients.
# reward_coeff: 0.5
# expectile_weight: 0.9
# value_coeff: 0.1
# consistency_coeff: 20.0
# advantage_scaling: 3.0
# pi_coeff: 0.5
# temporal_decay_coeff: 0.5
# # Target model.
# target_model_momentum: 0.995

53
lerobot/configs/robot/koch.yaml
View File

@@ -1,53 +0,0 @@
_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
robot_type: koch
calibration_dir: .cache/calibration/koch
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: null
leader_arms:
main:
_target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
port: /dev/tty.usbmodem58760430441
motors:
# name: (index, model)
shoulder_pan: [1, "xl330-m077"]
shoulder_lift: [2, "xl330-m077"]
elbow_flex: [3, "xl330-m077"]
wrist_flex: [4, "xl330-m077"]
wrist_roll: [5, "xl330-m077"]
gripper: [6, "xl330-m077"]
follower_arms:
main:
_target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
port: /dev/tty.usbmodem585A0083391
motors:
# name: (index, model)
shoulder_pan: [1, "xl430-w250"]
shoulder_lift: [2, "xl430-w250"]
elbow_flex: [3, "xl330-m288"]
wrist_flex: [4, "xl330-m288"]
wrist_roll: [5, "xl330-m288"]
gripper: [6, "xl330-m288"]
cameras:
laptop:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 0
fps: 30
width: 640
height: 480
phone:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 1
fps: 30
width: 640
height: 480
# ~ Koch specific settings ~
# Sets the leader arm in torque mode with the gripper motor set to this angle. This makes it possible
# to squeeze the gripper and have it spring back to an open position on its own.
gripper_open_degree: 35.156

63
lerobot/configs/robot/so100.yaml
View File

@@ -1,63 +0,0 @@
# [SO-100 robot arm](https://github.com/TheRobotStudio/SO-ARM100)
# Requires installing extras packages
# With pip: `pip install -e ".[feetech]"`
# With poetry: `poetry install --sync --extras "feetech"`
# See [tutorial](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md)
_target_: lerobot.common.robot_devices.robots.manipulator.ManipulatorRobot
robot_type: so100
calibration_dir: .cache/calibration/so100
# `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
# Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
# the number of motors in your follower arms.
max_relative_target: null
joint_position_relative_bounds: null
# max: [100, 100, 100, 100, 100, 100]
# min: [-100, -100, -100, -100, -100, -100]
# max: [ 7.2158203e+01, 1.5398438e+02, 1.6075195e+02, 9.3251953e+01, 0., -1.4184397e-01]
# min: [-77.08008, 56.25, 60.55664, 19.511719, 0., -0.63829786]
# max: [ 35.06836 , 103.18359 , 127.61719 , 75.58594 , 0., 0.]
# min: [ -8.876953 , 63.808594 , 90.49805 , 49.48242 , 0., 0.]
leader_arms:
main:
_target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
port: /dev/tty.usbmodem58760433331
motors:
# name: (index, model)
shoulder_pan: [1, "sts3215"]
shoulder_lift: [2, "sts3215"]
elbow_flex: [3, "sts3215"]
wrist_flex: [4, "sts3215"]
wrist_roll: [5, "sts3215"]
gripper: [6, "sts3215"]
follower_arms:
main:
_target_: lerobot.common.robot_devices.motors.feetech.FeetechMotorsBus
port: /dev/tty.usbmodem58760431631
motors:
# name: (index, model)
shoulder_pan: [1, "sts3215"]
shoulder_lift: [2, "sts3215"]
elbow_flex: [3, "sts3215"]
wrist_flex: [4, "sts3215"]
wrist_roll: [5, "sts3215"]
gripper: [6, "sts3215"]
cameras:
front:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 1
fps: 30
width: 640
height: 480
side:
_target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
camera_index: 0
fps: 30
width: 640
height: 480