v0.6 (#18)
* Sampling over both env and denoising steps in DPPO updates (#13) * sample one from each chain * full random sampling * Add Proficient Human (PH) Configs and Pipeline (#16) * fix missing cfg * add ph config * fix how terminated flags are added to buffer in ibrl * add ph config * offline calql for 1M gradient updates * bug fix: number of calql online gradient steps is the number of new transitions collected * add sample config for DPPO with ta=1 * Sampling over both env and denoising steps in DPPO updates (#13) * sample one from each chain * full random sampling * fix diffusion loss when predicting initial noise * fix dppo inds * fix typo * remove print statement --------- Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu> Co-authored-by: allenzren <allen.ren@princeton.edu> * update robomimic configs * better calql formulation * optimize calql and ibrl training * optimize data transfer in ppo agents * add kitchen configs * re-organize config folders, rerun calql and rlpd * add scratch gym locomotion configs * add kitchen installation dependencies * use truncated for termination in furniture env * update furniture and gym configs * update README and dependencies with kitchen * add url for new data and checkpoints * update demo RL configs * update batch sizes for furniture unet configs * raise error about dropout in residual mlp * fix observation bug in bc loss --------- Co-authored-by: Justin Lidard <60638575+jlidard@users.noreply.github.com> Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu>
This commit is contained in:
Allen Z. Ren
co-authored by
Justin M. Lidard
Justin Lidard
Justin M. Lidard
Justin Lidard
parent
7b10df690d
commit
dc8e0c9edc
@@ -0,0 +1,114 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_rlpd_agent.TrainRLPDAgent
|
||||
|
||||
name: ${env_name}_rlpd_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/train.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: can
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 1
|
||||
name: ${env_name}
|
||||
max_episode_steps: 300
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 1
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: rlpd-${env_name}
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000000
|
||||
n_steps: 1
|
||||
gamma: 0.99
|
||||
actor_lr: 1e-4
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
critic_lr: 1e-4
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
save_model_freq: 100000
|
||||
val_freq: 10000
|
||||
render:
|
||||
freq: 10000
|
||||
num: 0
|
||||
log_freq: 200
|
||||
# RLPD specific
|
||||
batch_size: 256
|
||||
target_ema_rate: 0.01
|
||||
scale_reward_factor: 1
|
||||
critic_num_update: 3
|
||||
buffer_size: 400000
|
||||
n_eval_episode: 40
|
||||
n_explore_steps: 0
|
||||
target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
|
||||
init_temperature: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
||||
randn_clip_value: 10
|
||||
backup_entropy: True
|
||||
n_critics: 5
|
||||
tanh_output: True
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
tanh_output: False # squash after sampling instead
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
std_max: 7.3891
|
||||
std_min: 0.0067
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: ReLU
|
||||
use_layernorm: True
|
||||
double_q: False # use ensemble
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
device: ${device}
|
||||
|
||||
offline_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
||||
dataset_path: ${offline_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,114 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_rlpd_agent.TrainRLPDAgent
|
||||
|
||||
name: ${env_name}_rlpd_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-ph/normalization.npz
|
||||
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-ph/train.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: can
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 1
|
||||
name: ${env_name}
|
||||
max_episode_steps: 300
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 1
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: rlpd-${env_name}
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000000
|
||||
n_steps: 1
|
||||
gamma: 0.99
|
||||
actor_lr: 1e-4
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
critic_lr: 1e-4
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
save_model_freq: 100000
|
||||
val_freq: 10000
|
||||
render:
|
||||
freq: 10000
|
||||
num: 0
|
||||
log_freq: 200
|
||||
# RLPD specific
|
||||
batch_size: 256
|
||||
target_ema_rate: 0.01
|
||||
scale_reward_factor: 1
|
||||
critic_num_update: 3
|
||||
buffer_size: 400000
|
||||
n_eval_episode: 40
|
||||
n_explore_steps: 0
|
||||
target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
|
||||
init_temperature: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
||||
randn_clip_value: 10
|
||||
backup_entropy: True
|
||||
n_critics: 5
|
||||
tanh_output: True
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
tanh_output: False # squash after sampling instead
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
std_max: 7.3891
|
||||
std_min: 0.0067
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: ReLU
|
||||
use_layernorm: True
|
||||
double_q: False # use ensemble
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
device: ${device}
|
||||
|
||||
offline_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
||||
dataset_path: ${offline_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,114 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_rlpd_agent.TrainRLPDAgent
|
||||
|
||||
name: ${env_name}_rlpd_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/train.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 1
|
||||
name: ${env_name}
|
||||
max_episode_steps: 400
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 1
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: rlpd-${env_name}
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000000
|
||||
n_steps: 1
|
||||
gamma: 0.99
|
||||
actor_lr: 1e-4
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
critic_lr: 1e-4
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
save_model_freq: 100000
|
||||
val_freq: 10000
|
||||
render:
|
||||
freq: 10000
|
||||
num: 0
|
||||
log_freq: 200
|
||||
# RLPD specific
|
||||
batch_size: 256
|
||||
target_ema_rate: 0.01
|
||||
scale_reward_factor: 1
|
||||
critic_num_update: 3
|
||||
buffer_size: 400000
|
||||
n_eval_episode: 40
|
||||
n_explore_steps: 0
|
||||
target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
|
||||
init_temperature: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
||||
randn_clip_value: 10
|
||||
backup_entropy: True
|
||||
n_critics: 5
|
||||
tanh_output: True
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
tanh_output: False # squash after sampling instead
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
std_max: 7.3891
|
||||
std_min: 0.0067
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: ReLU
|
||||
use_layernorm: True
|
||||
double_q: False # use ensemble
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
device: ${device}
|
||||
|
||||
offline_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
||||
dataset_path: ${offline_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,114 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_rlpd_agent.TrainRLPDAgent
|
||||
|
||||
name: ${env_name}_rlpd_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-ph/normalization.npz
|
||||
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-ph/train.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 1
|
||||
name: ${env_name}
|
||||
max_episode_steps: 400
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 1
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: rlpd-${env_name}
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000000
|
||||
n_steps: 1
|
||||
gamma: 0.99
|
||||
actor_lr: 1e-4
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
critic_lr: 1e-4
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
save_model_freq: 100000
|
||||
val_freq: 10000
|
||||
render:
|
||||
freq: 10000
|
||||
num: 0
|
||||
log_freq: 200
|
||||
# RLPD specific
|
||||
batch_size: 256
|
||||
target_ema_rate: 0.01
|
||||
scale_reward_factor: 1
|
||||
critic_num_update: 3
|
||||
buffer_size: 400000
|
||||
n_eval_episode: 40
|
||||
n_explore_steps: 0
|
||||
target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
|
||||
init_temperature: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
||||
randn_clip_value: 10
|
||||
backup_entropy: True
|
||||
n_critics: 5
|
||||
tanh_output: True
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
tanh_output: False # squash after sampling instead
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
std_max: 7.3891
|
||||
std_min: 0.0067
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: ReLU
|
||||
use_layernorm: True
|
||||
double_q: False # use ensemble
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
device: ${device}
|
||||
|
||||
offline_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
||||
dataset_path: ${offline_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
Reference in New Issue
Block a user