lerobot/lerobot/configs/policy/sac_real.yaml

121 lines
3.3 KiB
YAML

# @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