Release DiscoDemo RL

README.md

@@ -0,0 +1,48 @@
+---
+license: apache-2.0
+tags:
+- robotics
+- reinforcement-learning
+- discodemo
+- franka
+---
+
+# DiscoDemo-Stage1_RL-FMB_Round-alpha0
+
+The **data generator** of [DiscoDemo](https://davian-robotics.github.io/DiscoDemo/) for the **FMB-Round** task (insert a cylindrical peg into its hole on the FMB board): a state-based
+reinforcement-learning policy trained in simulation — the P-RFCL baseline generator (α = 0: the same reverse-curriculum RL without the diversity reward).
+Rolling it out generates demonstrations such as those in the Stage2 dataset below.
+
+| Stage | Repository |
+|---|---|
+| Generator (RL policy) | [`DAVIAN-Robotics/DiscoDemo-Stage1_RL-FMB_Round-alpha0`](https://huggingface.co/DAVIAN-Robotics/DiscoDemo-Stage1_RL-FMB_Round-alpha0) (this repository) |
+| Generated dataset | [`DAVIAN-Robotics/DiscoDemo-Stage2_GenData-FMB_Round-alpha0`](https://huggingface.co/datasets/DAVIAN-Robotics/DiscoDemo-Stage2_GenData-FMB_Round-alpha0) |
+| Imitation policy (π0.5) | [`DAVIAN-Robotics/DiscoDemo-Stage3_SFT-FMB_Round-alpha0`](https://huggingface.co/DAVIAN-Robotics/DiscoDemo-Stage3_SFT-FMB_Round-alpha0) |
+
+## Files
+
+| 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. |
+| `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.
+
+
+Environment steps at this checkpoint: 46.5 M.
+
+## Usage
+
+Loading requires the DiscoDemo training code: https://github.com/DAVIAN-Robotics/DiscoDemo.
+
+## Citation
+
+```bibtex
+@article{park2026discodemo,
+  title   = {DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning},
+  author  = {Park, Minho and Kim, Kinam and Kim, Donghu and Lee, Byungkun and Hwang, Dongyoon and Shin, Yongjae and Hyung, Junha and Lee, Hojoon and Choo, Jaegul},
+  journal = {arXiv preprint},
+  year    = {2026}
+}
+```

actor.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db33561e501f3161bb2ec2257938c3848be9370da23ae19ebc878e368a102deb
+size 1109185

config.yaml

@@ -0,0 +1,473 @@
+agent:
+  agent_type: flashSAC
+  seed: 0
+  device_type: cuda
+  buffer_max_length: 100000000
+  buffer_min_length: 5000
+  buffer_device_type: cuda
+  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
+  learning_rate_warmup_rate: 1.0e-06
+  learning_rate_warmup_step: 50
+  learning_rate_decay_rate: 1.0
+  learning_rate_decay_step: 50000000
+  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
+  temp_initial_value: 0.01
+  temp_target_sigma: 0.15
+  temp_target_entropy: null
+  gamma: 0.99
+  n_step: 8
+  use_compile: true
+  compile_mode: auto
+  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
+    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
+  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
+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
+    - panda_link7
+    - panda_hand
+    - panda_leftfinger
+    - panda_rightfinger
+    scene_bodies:
+    - 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
+  obj_press:
+    weight: 1.0e-06
+    deadband_axial: 5.0
+    deadband_lateral: 15.0
+    sat: 100.0
+    surfaces:
+    - board
+save_video: true
+video_camera: exo_wrist
+disable_train_cameras: true
+control_decimation: 6
+jerk_limited_action:
+  enabled: true
+  max_joint_velocity:
+  - 0.22215
+  - 0.51715
+  - 0.19244
+  - 0.66218
+  - 0.31884
+  - 0.41767
+  - 0.40726
+  max_joint_acceleration:
+  - 0.46616
+  - 0.95012
+  - 0.46407
+  - 1.28545
+  - 1.25775
+  - 1.18572
+  - 1.33469
+  max_joint_jerk:
+  - 2.65668
+  - 4.40304
+  - 2.82091
+  - 5.53682
+  - 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

normalizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95db1c12b4e976b513cf21aaabb8b35f4aa324391498d562a5271b97ca3a058d
+size 3002