Config and normalizer in the format of the current DiscoDemo code

README.md

@@ -24,7 +24,7 @@ Rolling it out generates demonstrations such as those in the Stage2 dataset belo
 | File | Content |
 |---|---|
 | `actor.pt` | Actor network weights (`network_state_dict`). Optimizer state is not included. |
-| `normalizer.pt` | Running statistics used to normalize the policy input. |
+| `normalizer.pt` | Observation bounds used to normalize the policy input. |
 | `config.yaml` | Full training configuration (environment, curriculum, agent). |
 
 Only what is needed to roll the policy out is released; the critic and other training-only state are omitted.

config.yaml

@@ -1,3 +1,22 @@
+project_name: DiscoDemo
+entity_name: null
+group_name: fmb_round
+exp_name: prfcl-fmb_round
+seed: 0
+logger_type: wandb
+save_path: libs/FlashSAC/models/fmb_round/prfcl-fmb_round/seed0-TIMESTAMP
+agent_load_path: null
+num_env_steps: 50000000
+num_train_envs: 2048
+updates_per_interaction_step: 128
+gamma: 0.99
+n_step: 8
+env:
+  workcell: libs/real2sim/assets/workcell.json
+  task: libs/real2sim/real2sim/tasks/fmb_round.py
+  env_version: fr3-fmb_round
+  gripper_lock: closed
+success_hold_steps: 20
 agent:
   agent_type: flashSAC
   seed: 0
@@ -8,9 +27,7 @@ agent:
   sample_batch_size: 2048
   normalize_reward: true
   normalized_G_max: 5.0
-  reward_grmax_clamp: 0.0
   asymmetric_observation: false
-  store_ep_uid: false
   learning_rate_init: 0.0003
   learning_rate_peak: 0.0003
   learning_rate_end: 0.00015
@@ -21,14 +38,12 @@ agent:
   actor_num_blocks: 2
   actor_hidden_dim: 128
   actor_bc_alpha: 0.0
-  actor_action_magnitude_coef: 0.0
   actor_noise_zeta_mu: 2.0
   actor_noise_zeta_max: 16
   actor_update_period: 2
   critic_num_blocks: 2
   critic_hidden_dim: 256
   critic_num_bins: 101
-  critic_value_range: 5
   critic_min_v: -5
   critic_max_v: 5
   critic_target_update_tau: 0.01
@@ -42,211 +57,32 @@ agent:
   use_amp: true
   load_optimizer: true
   load_reward_normalizer: true
-  normalize_inputs: true
-  normalize_clamp: true
-  normalize_inputs_mode: bounds
-  normalize_inputs_skip_keys: []
-env:
-  env_type: robolab
-  env_name: Fr3PegInsertRound
-  num_env_steps: 50000000
-  seed: 0
-  num_train_envs: 2048
-  num_eval_envs: 0
-  num_record_envs: 1
-  rescale_action: true
-  max_episode_steps: null
-  robot_variant: fr3
-  sysid_workcell: libs/real2sim/assets/workcell.json
-  canonical_env_version: fr3-peg-insert-round-v3
-  gripper_close_width_m: null
-  gripper_stiffness: null
-  gripper_damping: null
-  gripper_delay: true
-  gripper_delay_obs: true
-  gripper_delay_obs_mode: elapsed
-  gripper_lock: closed
-  action_repeat: 1
-  task: libs/real2sim/real2sim/tasks/fr3_peg_insert_round.py
-  reward_mode: sparse
-  headless: true
-seed: 0
-load_replay_buffer_on_resume: true
-agent_load_skip_input_layer: false
-num_env_steps: 50000000
-num_train_envs: 2048
-num_eval_envs: 0
-num_record_envs: 1
-num_eval_episodes: 1
-num_record_episodes: 1
-gamma: 0.99
-n_step: 8
-num_interaction_steps: 24414.0625
-updates_per_interaction_step: 128
-num_update_steps: 3125000.0
-evaluation_per_interaction_step: 244
-metrics_per_interaction_step: 244
-recording_per_interaction_step: 813
-logging_per_interaction_step: 24
-save_checkpoint_per_interaction_step: 244
-save_buffer_per_interaction_step: 244
-keep_only_latest_periodic_buffer: true
-eval_random_seed: 12345
-reconfigure_per_interaction_step: null
 rfcl:
-  num_demos: 20
-  skip_failed: true
-  shuffle_demos: false
-  reward_mode: sparse
   demo_ratio: 0.5
   truncate_after_success: 50
-  skip_reverse_curriculum: false
   reverse:
-    curriculum_method: per_demo
     reverse_step_size: 16
-    per_env_buffer_size: 3
     advance_threshold: 0.9
-    start_step_sampler: multienv-zvf
-    demo_horizon_to_max_steps_ratio: 1
-    verbose: 1
-    num_train_envs: 20
-    updates_per_interaction_step: null
-    curriculum_keyframes:
-      enabled: false
-      path: null
-      remap_dense_resume: false
-    multi_env: null
+    frontier_window: 10
     minimum_episode_steps: 30
-    zvf_hi: 2.0
-    zvf_lo: 0.0
-    zvf_window_K: 10
-    zvf_sample_weight: geometric
-    zvf_mask: false
-  forward:
-    enabled: true
-    num_seeds: 1000
-    rho: 1.0
-    nu: 0.2
-    score_fn: success_once_score
-    score_transform: rankmin
-    score_temperature: 0.1
-    staleness_transform: rankmin
-    staleness_temperature: 0.1
-    staleness_coef: 0.1
-  transition:
-    enabled: true
-    solved_frac_threshold: 0.5
-    save_checkpoint_and_buffer: true
-    force_forward_at_is: 0
+  solved_frac_threshold: 0.5
+  demo_path: data/demo_banks/fmb_round/human_demos.h5
   init_pose: pregrasp:0.03,15
-  init_pose_source: onthefly
-  init_pose_root: null
-  init_pose_start_row: null
-  demo_path: null
-  demo_bank: capx20-round-v3
-resume_forward: false
 eval_worker:
-  enabled: true
-  measure_safety: false
   num_envs: 50
-  video_max_steps: 0
-  video_speed: 2.0
-  video_scale: 0.5
-  video_quality: 5
-  frontier_eval:
-    enabled: true
-    num_demos: 8
-    envs_per_demo: 6
-    frame_offset: 0
-    video: true
-    exclude_terminal_frontiers: false
-    terminal_slack_steps: -1
-eval_success_window_size: 0
-max_episode_steps: 0
-success_hold_steps: 20
-success_reward_mode: oneshot
-success_as_truncation: false
-timeout_as_termination: true
-distractor_obs: true
-settle_from_reverse: true
-adaptive_horizon:
-  enabled: false
-  multiplier: 2.0
-  min_steps: 50
-  min_episodes: 10
-max_relative_action_rate_k: null
-max_relative_action_rate_abs:
-- 0.0111075
-- 0.0258575
-- 0.009622
-- 0.033109
-- 0.015942
-- 0.0208835
-- 0.020363
 skill:
-  z_dim: 0
-  z_unit: false
-  z_norm_match: false
-  alpha: 0.0
-  phi_space: object
   phi_hidden:
   - 256
   - 256
   phi_lr: 0.0001
   dual_slack: 0.001
-  alpha_mode: raw
-  budget_warmup_updates: 500
-  budget_freeze_after: 0
   lam_init: 1.0
-  gate_capacity: 1048576
-  skill_type: metra
-  z_discrete: false
-  n_skills: 5
-  disc_hidden:
-  - 256
-  - 256
-  disc_lr: 0.0001
-partial_credit:
-  coef: 0.0
-living_cost:
-  coef: 0.0
+  alpha: 0.0
+  z_dim: 0
 safety_penalty:
-  enabled: true
-  apply_in_reverse: true
-  cap: 0.1
-  qvel:
-    coef: 0.0
-    deadband: 0.12
-    sat: 4.44
-  qacc:
-    coef: 0.0
-    deadband: 14.0
-    sat: 27183
-  action_smoothness:
-    coef: 0.0
-    deadband: 0.1
-    sat: 0.265
-  rel_action:
-    coef: 0.0
-    deadband: 0.5
-    norm: 2
-    sat: 1.0
-  gripper_flip:
-    coef: 0.0
-    deadband: 0.0
-    sat: 1.0
-  eef_lin_vel:
-    weight: 0.0
-    deadband: 0.18
-    sat: 1.5
   illegal_contact:
     weight: 0.0001
     sat: 100
-    allowed_pairs:
-    - - panda_leftfinger
-      - peg
-    - - panda_rightfinger
-      - peg
     robot_bodies:
     - panda_link5
     - panda_link6
@@ -258,34 +94,11 @@ safety_penalty:
     - peg
     - board
     - table
-  obj_lin_vel:
-    weight: 0.0
-    deadband: 0.05
-    sat: 2.04
-  obj_ang_vel:
-    weight: 0.0
-    deadband: 0.5
-    sat: 28.8
-  obj_acc:
-    weight: 0.0
-    deadband: 0.0
-    sat: 29.6
-  obj_impact:
-    weight: 0.0
-    deadband: 1.0
-    sat: 553
-  other_vel:
-    weight: 0.0
-    deadband: 0.05
-    sat: 2.18
-  other_disp:
-    weight: 0.0
-    deadband: 0.02
-    sat: 1.98
-  other_impact:
-    weight: 0.0
-    deadband: 1.0
-    sat: 414
+    allowed_pairs:
+    - - panda_leftfinger
+      - peg
+    - - panda_rightfinger
+      - peg
   obj_press:
     weight: 1.0e-06
     deadband_axial: 5.0
@@ -293,12 +106,27 @@ safety_penalty:
     sat: 100.0
     surfaces:
     - board
-save_video: true
-video_camera: exo_wrist
-disable_train_cameras: true
-control_decimation: 6
+obs_bounds:
+  object_z_m:
+  - 0.0
+  - 0.6
+  eef_box_m:
+    x:
+    - 0.2
+    - 0.85
+    'y':
+    - -0.45
+    - 0.45
+    z:
+    - 0.0
+    - 0.7
+  object_xy:
+  - - 0.25
+    - 0.65
+  - - -0.25
+    - 0.15
+task_name: fmb_round
 jerk_limited_action:
-  enabled: true
   max_joint_velocity:
   - 0.22215
   - 0.51715
@@ -323,151 +151,4 @@ jerk_limited_action:
   - 9.90014
   - 6.67391
   - 10.55754
-reset_reverse_curriculum_on_resume: false
-obs_bounds:
-  object_xy:
-  - - 0.25
-    - 0.65
-  - - -0.25
-    - 0.15
-  object_xy_margin_m: 0.1
-  object_z_m:
-  - 0.0
-  - 0.6
-  eef_box_m:
-    x:
-    - 0.2
-    - 0.85
-    y:
-    - -0.45
-    - 0.45
-    z:
-    - 0.0
-    - 0.7
-  joints: physical
-  resolved:
-    terms:
-    - - target0_pos
-      - 3
-    - - target0_quat
-      - 4
-    - - target0_tcp
-      - 3
-    - - container_pos
-      - 3
-    - - container_quat
-      - 4
-    - - tcp_to_container
-      - 3
-    - - arm_joint_pos
-      - 7
-    - - gripper_pos
-      - 1
-    - - eef_position
-      - 3
-    - - eef_orientation
-      - 4
-    - - gripper_delay
-      - 2
-    - - prev_action
-      - 8
-    low:
-    - 0.25
-    - -0.25
-    - 0.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -0.6000000238418579
-    - -0.699999988079071
-    - -0.699999988079071
-    - 0.25
-    - -0.25
-    - 0.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -0.6000000238418579
-    - -0.699999988079071
-    - -0.699999988079071
-    - -2.8973002433776855
-    - -1.7627999782562256
-    - -2.8973000049591064
-    - -3.0717999935150146
-    - -2.8973000049591064
-    - -0.017499923706054688
-    - -3.6826982498168945
-    - 0.0
-    - 0.20000000298023224
-    - -0.44999998807907104
-    - 0.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - 0.0
-    - 0.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    - -1.0
-    high:
-    - 0.6499999761581421
-    - 0.15000000596046448
-    - 0.6000000238418579
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 0.44999998807907104
-    - 0.6000000238418579
-    - 0.6000000238418579
-    - 0.6499999761581421
-    - 0.15000000596046448
-    - 0.6000000238418579
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 0.44999998807907104
-    - 0.6000000238418579
-    - 0.6000000238418579
-    - 2.8977999687194824
-    - 1.7627999782562256
-    - 2.8973000049591064
-    - -0.06979990005493164
-    - 2.8973000049591064
-    - 3.752500057220459
-    - 2.1119017601013184
-    - 1.0
-    - 0.8500000238418579
-    - 0.44999998807907104
-    - 0.699999988079071
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-    - 1.0
-max_relative_action_rate:
-- 0.0111075
-- 0.0258575
-- 0.009622
-- 0.033109
-- 0.015942
-- 0.0208835
-- 0.020363
+method_name: prfcl

normalizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:95db1c12b4e976b513cf21aaabb8b35f4aa324391498d562a5271b97ca3a058d
-size 3002
+oid sha256:7b237a781e41219873d94bb327cb71697cf9f4b462c65ff98b96a4a92b9c2253
+size 2364