Add files using upload-large-folder tool

.gitattributes

@@ -61,3 +61,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 
 # 3D point clouds
 *.ply filter=lfs diff=lfs merge=lfs -text
+
+# Archives
+*.zip filter=lfs diff=lfs merge=lfs -text

README.md

@@ -28,13 +28,24 @@ Dynamics** (CoRL 2026).
 
 ```
 data/
-├── phystwin/different_types/<case>/    14 cases, redistributed (see Licenses)
-└── physcore/different_types/<case>/    12 cases, recorded by us
-gaussian_output/physcore/<case>/        static 3DGS scenes for the 12 physcore cases
-checkpoints/                            MfM_checkpoint.pt, RfD_checkpoint.pt
-configs/                                one YAML per pipeline script
+├── phystwin.zip                 14 cases, redistributed (see Licenses)
+└── physcore.zip                 12 cases, recorded by us
+gaussian_output/physcore.zip     static 3DGS scenes for the 12 physcore cases
+checkpoints/                     MfM_checkpoint.pt, RfD_checkpoint.pt
+configs/                         one YAML per pipeline script
 ```
 
+The recordings are shipped as archives to keep the repository to a handful of files.
+Unzip each one in place; every archive expands into a directory of the same name:
+
+```bash
+cd data            && unzip phystwin.zip && unzip physcore.zip && cd ..
+cd gaussian_output && unzip physcore.zip && cd ..
+```
+
+which gives `data/phystwin/different_types/<case>/`,
+`data/physcore/different_types/<case>/` and `gaussian_output/physcore/<case>/`.
+
 Each of the 26 cases is one recording from **3 calibrated RGB-D cameras**, time-aligned:
 
 | path | content |
@@ -50,7 +61,8 @@ Each of the 26 cases is one recording from **3 calibrated RGB-D cameras**, time-
 
 ## Where to start
 
-Copy the contents of this repository into the root of a PhysCoRe checkout, so that
+Unzip the three archives as above, then copy the contents of this repository into the
+root of a PhysCoRe checkout, so that
 `data/`, `gaussian_output/`, `checkpoints/` and `configs/` sit beside the pipeline
 scripts. The code repository's README walks through the stages in order and names the
 config each one reads.

configs/augment_recipe.yaml

@@ -0,0 +1,170 @@
+# Augmentation recipe for datagen/augment/run_recipe.py.
+#
+# Each group is self-contained: material sampling, rollout physics, rendering,
+# seed, and output count all live here.
+
+sources_root: data_episodes/phystwin
+output_root: data_episodes_augmented/episodes
+viz_root: data_episodes_augmented/video
+device: cuda
+viz: true
+
+groups:
+  - name: elastic
+    mode: elastic
+    num_episodes: 70
+    seed_base: 42
+    material:
+      log_E:
+        mean_range: [6.0, 9.0]
+        std_range: [0.5, 1.5]
+        clamp: [5.0, 11.0]
+      nu:
+        mean_range: [0.15, 0.35]
+        std_range: [0.05, 0.15]
+        clamp: [0.05, 0.45]
+      noise:
+        freq_range: [0.5, 1.6]
+        octaves: 4
+        persistence: 0.55
+        lacunarity: 2.0
+    render:
+      max_observed_points: 0
+      particle_subset: shell
+      splat_radius: 1
+      depth_noise_std: 0.002
+      depth_noise_relative: 0.001
+      depth_dropout: 0.01
+      near_depth: 0.01
+      far_depth: 60.0
+    ground:
+      friction: 0.5
+      elasticity: 0.0
+      clearance_cells: 0.0
+    dataset:
+      rollout_steps: 25
+      real_world_domain_center: [0.5, 0.5, 0.2]
+    rollout:
+      use_batched_mpm: true
+      use_episode_sim_dt: true
+      default_dt: 0.000666
+      default_steps_per_frame: 2
+      num_grids: 32
+      clip_bound: 0.5
+      gravity: [0.0, 0.0, -9.8]
+      damping: 0.999
+      rho: 100.0
+      ground_height: 0.02
+      ground_surface: coulomb
+      ground_elasticity: 0.0
+      ground_friction: 0.5
+      ground_tangent_damping: 0.25
+      ground_static_velocity_threshold: 0.02
+      use_explicit_rigid_contact: true
+      use_episode_rigid_contact: true
+      rigid_surface_fps: 256
+      rigid_friction: 0.5
+      rigid_contact_margin: 0.02
+      rigid_velocity_blend: 0.9
+      rigid_stickiness: 1.0
+      rigid_strict_nonpenetration: true
+      rigid_projection_iterations: 2
+      rigid_projection_margin: 5.0e-05
+      elasticity_model: CorotatedElasticity
+      plasticity_model: IdentityPlasticity
+      manipulation_controller_grid_contact_radius: 0.08
+      manipulation_controller_grid_velocity_blend: 1.0
+      max_delta_velocity_norm: 0.05
+      correction_target_velocity_gain: 100.0
+      correction_target_velocity_damping: 1.2
+      correction_target_velocity_integral_gain: 40.0
+      correction:
+        mode: spring
+        dt_scale: 0.25
+        steps_per_frame_scale: 4
+        velocity_cap: 5.0
+        omega_n: 150.0
+        zeta: 2.5
+      apply_force_every_substep: true
+      max_force_norm: 5000.0
+
+  - name: plastic
+    mode: plastic
+    num_episodes: 10
+    seed_base: 44
+    material:
+      log_E:
+        mean_range: [10.0, 10.5]
+        std_range: [0.0, 0.0]
+        clamp: [10.0, 10.5]
+      nu:
+        mean_range: [0.30, 0.38]
+        std_range: [0.0, 0.0]
+        clamp: [0.30, 0.40]
+      noise:
+        freq_range: [1.6, 2.4]
+        octaves: 4
+        persistence: 0.55
+        lacunarity: 2.0
+    render:
+      max_observed_points: 0
+      particle_subset: shell
+      splat_radius: 1
+      depth_noise_std: 0.002
+      depth_noise_relative: 0.001
+      depth_dropout: 0.01
+      near_depth: 0.01
+      far_depth: 60.0
+    ground:
+      friction: 0.5
+      elasticity: 0.0
+      clearance_cells: 0.0
+    dataset:
+      rollout_steps: 25
+      real_world_domain_center: [0.5, 0.5, 0.2]
+    rollout:
+      use_batched_mpm: true
+      use_episode_sim_dt: true
+      default_dt: 0.000666
+      default_steps_per_frame: 2
+      num_grids: 32
+      clip_bound: 0.5
+      gravity: [0.0, 0.0, -9.8]
+      damping: 0.999
+      rho: 100.0
+      ground_height: 0.02
+      ground_surface: coulomb
+      ground_elasticity: 0.0
+      ground_friction: 0.5
+      ground_tangent_damping: 0.25
+      ground_static_velocity_threshold: 0.02
+      use_explicit_rigid_contact: true
+      use_episode_rigid_contact: true
+      rigid_surface_fps: 256
+      rigid_friction: 0.5
+      rigid_contact_margin: 0.02
+      rigid_velocity_blend: 0.9
+      rigid_stickiness: 1.0
+      rigid_strict_nonpenetration: true
+      rigid_projection_iterations: 2
+      rigid_projection_margin: 5.0e-05
+      elasticity_model: CorotatedElasticity
+      plasticity_model: VonMisesPlasticity
+      plasticity_kwargs:
+        VonMisesPlasticity:
+          sigma_y: 200.0
+      manipulation_controller_grid_contact_radius: 0.08
+      manipulation_controller_grid_velocity_blend: 1.0
+      max_delta_velocity_norm: 0.05
+      correction_target_velocity_gain: 100.0
+      correction_target_velocity_damping: 1.2
+      correction_target_velocity_integral_gain: 40.0
+      correction:
+        mode: spring
+        dt_scale: 0.25
+        steps_per_frame_scale: 4
+        velocity_cap: 5.0
+        omega_n: 150.0
+        zeta: 2.5
+      apply_force_every_substep: true
+      max_force_norm: 5000.0

configs/render_MfM_confidence_3dgs.yaml

@@ -0,0 +1,94 @@
+# MfM confidence -> dynamic 3DGS video. Owned by render_MfM_confidence_3dgs.py.
+# Needs `validate_MfM.py --save-rollout` on the same checkpoint and episode.
+
+render:
+  checkpoint: checkpoints/MfM_checkpoint.pt   # must be the one that produced traj_path
+  episode_root: data_episodes/physcore/<case_name>/episode_0000
+  gs_dir: gaussian_output/physcore/<case_name>/init=pcd_iso=True_ldepth=0.001_lnormal=0.0_laniso_0.0_lseg=1.0
+  traj_path: outputs/MfM/val/YYYYMMDD_hhmm/validation_episode_0000_trajectory.pt
+  out_dir: outputs/MfM_conf_overlay/<case_name>_YYYYMMDD_hhmm   # IS the run dir; stamp by hand
+  name: null
+  device: cuda
+  mem_fraction: 0.0
+
+  overlay_on_rgb: true
+  rgb_root: null
+  overlay_alpha_gain: 1.0
+  overlay_bg_fade: 0.4
+  overlay_front_opacity: 0.95
+
+  channels: null
+  cmap: viridis
+  norm_q: 0.99               # used only when norm_hi is null
+  norm_hi: 2.0               # tune per case; pin it to compare runs
+  midpoint_only: false
+  running_max: false
+  warmup_frames: 30          # settles the recurrent state; 0 = off, and the video opens bright
+
+  iteration: 10000
+  opacity_threshold: 0.01
+  bg_dist_threshold: 0.03
+  knn_k: 16                  # also drives the warp weights, so this is not a color-only knob
+  warp_chunk: 4096
+
+  fps: 30
+  video_frames: null
+  save_pngs: true
+  metadata_json: null
+  flip_z: null
+
+# Must match what validate_MfM.py ran with; defines the shift undone to put the
+# trajectory back in the Gaussians' frame.
+dataset:
+  real_world_domain_center: [0.5, 0.5, 0.2]
+  observation_views: all
+  observation_view_index: 0
+
+# Refiner architecture. Must match the checkpoint; the model is rebuilt from
+# this block before the weights load.
+model:
+  hidden: 128
+  graph_unet_k: 8
+  graph_unet_pool: fps
+  graph_unet_ratios:
+  - 1.0
+  - 0.5
+  - 0.25
+  edge_feature: canonical_length_only
+  edge_direction: false
+  fixed_tracked_mask: false
+  use_persistent_tracks: true
+  persistent_track_use_motion_valid: true
+  tracked_disp_mode: canonical
+  control_disp_mode: absolute
+  max_controls: 4
+  tracked_disp_scale: 33.0
+  tracked_mask_scale: 1.0
+  control_vec_scale: 1.0
+  control_disp_scale: 33.0
+  canonical_bands: 4
+  tracked_bands: 16
+  attention_window: 9
+  temporal_conv: 9
+  initial_material_confidence_bias: 0.0
+  evidence_aggregation: true
+  evidence_tracked_weight: 8.0
+  evidence_control_weight: 4.0
+  evidence_control_radius: 0.04
+  log_E_min: 5.0
+  log_E_max: 11.0
+  nu_min: 0.05
+  nu_max: 0.45
+  graph_aggregation_chunks: 10
+
+train:
+  seed: 0                    # seeds material_guess below only
+  update_every: 10           # K — MfM refreshes every K frames and holds in between
+
+# This model ignores the initial guess, so it has no effect on the render.
+material_guess:
+  mode: uniform
+  log_E_range: [7.66, 7.66]
+  nu_range: [0.363, 0.363]
+  log_E_clamp_range: [5.0, 12.0]
+  nu_clamp_range: [0.05, 0.45]

configs/train_MfM.yaml

@@ -0,0 +1,99 @@
+# MfM (Material from Motion) training. Owned by train_MfM.py.
+# The names `refiner` and `Refiner`, here and in the code, both refer to MfM.
+
+dataset:
+  # Augmented episodes from run_recipe.py, relative to the repo root. The split
+  # below is the first 80% of each group for training, the last 20% for validation.
+  train_roots:
+  - data_episodes_augmented/episodes/*_elastic/episode_00[0-4][0-9]   # elastic 0000-0049
+  - data_episodes_augmented/episodes/*_elastic/episode_005[0-5]       # elastic 0050-0055
+  - data_episodes_augmented/episodes/*_plastic/episode_000[0-7]       # plastic 0000-0007
+  validation_roots:
+  - data_episodes_augmented/episodes/*_elastic/episode_005[6-9]       # elastic 0056-0059
+  - data_episodes_augmented/episodes/*_elastic/episode_006[0-9]       # elastic 0060-0069
+  - data_episodes_augmented/episodes/*_plastic/episode_000[8-9]       # plastic 0008-0009
+model:
+  hidden: 128
+  graph_unet_k: 8
+  graph_unet_pool: fps
+  graph_unet_ratios:
+  - 1.0
+  - 0.5
+  - 0.25
+  edge_feature: canonical_length_only
+  edge_direction: false
+  fixed_tracked_mask: false
+  use_persistent_tracks: true
+  persistent_track_use_motion_valid: true
+  tracked_disp_mode: canonical
+  control_disp_mode: absolute
+  max_controls: 4
+  tracked_disp_scale: 33.0
+  tracked_mask_scale: 1.0
+  control_vec_scale: 1.0
+  control_disp_scale: 33.0
+  canonical_bands: 4
+  tracked_bands: 16
+  attention_window: 9
+  temporal_conv: 9
+  initial_material_confidence_bias: 0.0
+  evidence_aggregation: true
+  evidence_tracked_weight: 8.0
+  evidence_control_weight: 4.0
+  evidence_control_radius: 0.04
+  log_E_min: 5.0
+  log_E_max: 11.0
+  nu_min: 0.05
+  nu_max: 0.45
+  graph_aggregation_chunks: 10
+train:
+  seed: 0
+  device: cuda
+  # Run directory, stamped by hand at launch: outputs/MfM/train/YYYYMMDD_hhmm.
+  output_dir: outputs/MfM/train/YYYYMMDD_hhmm
+  epochs: 2000
+  lr: 0.0001
+  final_lr: 0.0001
+  lr_warmup_epochs: 0
+  warmup_start_lr: 0.0001
+  lr_decay_start: 0
+  lr_decay_end: 0
+  weight_decay: 0.01
+  update_every: 10
+  tbptt_chunks: 4
+  batch_size: 4
+  batch_grouping: same_source
+  source_group_size: 64
+  supervision_fraction: 1.0
+  checkpoint_every: 10
+  best_checkpoint_metric: logE_shape
+  log_every_epochs: 1
+  save_latest_every_epoch: true
+  init_from: null
+  resume_from: null
+  validation_every: 10
+  shuffle: true
+  augment_z_rotation: true
+  tracking_dropout: 0.0
+  tracking_jitter_std: 0.007
+  use_source_motion_valid: true
+  temporal_crop_fraction: 0.0
+  cosine_decay: true
+loss:
+  material_weight: 1.0
+  material_confidence_log_weight: 0.2
+  plasticity_weight: 0.1
+material_guess:
+  mode: uniform
+  log_E_range:
+  - 7.66
+  - 7.66
+  nu_range:
+  - 0.363
+  - 0.363
+  log_E_clamp_range:
+  - 5.0
+  - 12.0
+  nu_clamp_range:
+  - 0.05
+  - 0.45

configs/train_RfD.yaml

@@ -0,0 +1,212 @@
+# RfD (Residual from Dynamics) training. Owned by train_RfD.py.
+# `gvc` = legacy RfD, `refiner` = MfM. validate_RfD.yaml needs the same output_dir.
+
+dataset:
+  train_roots:
+    - data_episodes/physcore/double_clift_cloth/episode_0000
+    - data_episodes/physcore/double_fold_rope/episode_0000
+    - data_episodes/physcore/double_squeeze_plastic/episode_0000
+    - data_episodes/physcore/double_stretch_bear/episode_0000
+    - data_episodes/physcore/double_stretch_bear_1/episode_0000
+    - data_episodes/physcore/double_stretch_plastic/episode_0000
+    - data_episodes/physcore/single_clift_cloth/episode_0000
+    - data_episodes/physcore/single_clift_rope/episode_0000
+    - data_episodes/physcore/single_lift_bear/episode_0000
+    - data_episodes/physcore/single_lift_cloth/episode_0000
+    - data_episodes/physcore/single_lift_rope/episode_0000
+    - data_episodes/physcore/single_push_rope/episode_0000
+  validation_roots:
+    - data_episodes/physcore/double_clift_cloth/episode_0000
+    - data_episodes/physcore/double_fold_rope/episode_0000
+    - data_episodes/physcore/double_squeeze_plastic/episode_0000
+    - data_episodes/physcore/double_stretch_bear/episode_0000
+    - data_episodes/physcore/double_stretch_bear_1/episode_0000
+    - data_episodes/physcore/double_stretch_plastic/episode_0000
+    - data_episodes/physcore/single_clift_cloth/episode_0000
+    - data_episodes/physcore/single_clift_rope/episode_0000
+    - data_episodes/physcore/single_lift_bear/episode_0000
+    - data_episodes/physcore/single_lift_cloth/episode_0000
+    - data_episodes/physcore/single_lift_rope/episode_0000
+    - data_episodes/physcore/single_push_rope/episode_0000
+  rollout_steps: 25
+  cache_size: 1
+  real_world_domain_center: [0.5, 0.5, 0.2]
+  observation_views: all
+  observation_view_index: 0
+
+# MPM rollout physics; per_sample_rollout_config below overrides it per case.
+rollout:
+  use_batched_mpm: true
+  use_episode_sim_dt: true
+  default_dt: 0.000666
+  default_steps_per_frame: 2
+  num_grids: 32
+  clip_bound: 0.5
+  gravity: [0.0, 0.0, -9.8]
+  damping: 0.999
+  rho: 100.0
+  ground_height: 0.02
+  ground_surface: coulomb
+  ground_elasticity: 0.0
+  ground_friction: 0.5
+  ground_tangent_damping: 0.25
+  ground_static_velocity_threshold: 0.02
+  use_explicit_rigid_contact: true
+  use_episode_rigid_contact: true
+  rigid_surface_fps: 256
+  rigid_friction: 0.5
+  rigid_contact_margin: 0.02
+  rigid_velocity_blend: 0.9
+  rigid_stickiness: 1.0
+  rigid_strict_nonpenetration: true
+  rigid_projection_iterations: 2
+  rigid_projection_margin: 5.0e-05
+  elasticity_model: CorotatedElasticity
+  plasticity_model: IdentityPlasticity
+  manipulation_controller_grid_contact_radius: 0.04
+  manipulation_controller_grid_velocity_blend: 1.0
+  max_delta_velocity_norm: 0.05
+  correction_target_velocity_gain: 100.0
+  correction_target_velocity_damping: 1.2
+  correction_target_velocity_integral_gain: 40.0
+  correction:
+    mode: spring
+    dt_scale: 0.25
+    steps_per_frame_scale: 4
+    velocity_cap: 5.0
+    omega_n: 150.0
+    zeta: 2.5
+  apply_force_every_substep: true
+  max_force_norm: 5000.0
+
+# Per-case overrides, keyed by case name. Radii are scene-dependent — retune for
+# your own objects.
+per_sample_rollout_config:
+  double_squeeze_plastic:
+    manipulation_controller_grid_contact_radius: 0.08
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  double_stretch_bear:
+    manipulation_controller_grid_contact_radius: 0.08
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  double_stretch_bear_1:
+    # Radius only — this case keeps the base IdentityPlasticity.
+    manipulation_controller_grid_contact_radius: 0.06
+  double_stretch_plastic:
+    manipulation_controller_grid_contact_radius: 0.08
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  single_lift_rope:
+    manipulation_controller_grid_contact_radius: 0.03
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  single_push_rope:
+    manipulation_controller_grid_contact_radius: 0.02
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+
+# Refiner (MfM) architecture. Must match gvc.refiner_checkpoint: the model is
+# built from this block and the weights loaded strictly.
+model:
+  spatial: graph_unet
+  temporal: conv_gru_fusion
+  hidden: 128
+  layers: 4
+  k: 8
+  graph_unet_k: 8
+  graph_unet_pool: fps
+  graph_unet_ratios: [1.0, 0.5, 0.25]
+  edge_feature: canonical_length_only
+  edge_direction: false
+  input_mode: observed_control
+  fixed_tracked_mask: false
+  use_persistent_tracks: true
+  persistent_track_use_motion_valid: true
+  tracked_disp_mode: canonical
+  control_disp_mode: absolute
+  max_controls: 4
+  observed_drop_canonical: false
+  tracked_disp_scale: 33.0
+  tracked_mask_scale: 1.0
+  control_vec_scale: 1.0
+  control_disp_scale: 33.0
+  coord_bands: 16
+  canonical_bands: 4
+  tracked_bands: 16
+  control_bands: 1
+  attention_window: 9
+  attention_heads: 4
+  temporal_conv: 9
+  use_material: false
+  use_correction: false
+  use_correction_mask: false
+  use_velocity: false
+  use_F: false
+  use_material_confidence: true
+  separate_confidence: true
+  detached_confidence: false
+  initial_material_confidence_bias: 0.0
+  aggregate_controls: true
+  evidence_aggregation: true
+  evidence_tracked_weight: 8.0
+  evidence_control_weight: 4.0
+  evidence_control_radius: 0.04
+  log_E_min: 5.0
+  log_E_max: 11.0
+  nu_min: 0.05
+  nu_max: 0.45
+  graph_aggregation_chunks: 10
+  use_decomposed_material: false
+
+# RfD, still spelled gvc here.
+gvc:
+  base_channels: 32
+  levels: 2
+  max_delta_v: 0.05
+  h: 10                    # corrector fires every h MPM substeps
+  warmup_fraction: 0.5     # first/second-half boundary; MfM freezes material there
+  bptt_frames: 5           # BPTT segment length in frames
+  refiner_checkpoint: checkpoints/MfM_checkpoint.pt
+
+train:
+  seed: 42
+  device: cuda
+  output_dir: outputs/RfD/YYYYMMDD_hhmm
+  epochs: 80
+  lr: 1.0e-4
+  final_lr: 1.0e-5
+  lr_decay_start: 30
+  lr_decay_end: 80
+  weight_decay: 0.0
+  grad_clip: 0.5
+  augment_z_rotation: true # per-(episode, epoch) z-rotation + 50% x-mirror flip
+
+  update_every: 10         # k = window size in camera frames
+  validation_every: 5
+  checkpoint_every: 5
+  plot_every: 1
+  viz_every: 20
+  viz_fps: 30
+  viz_max_controls: 16
+
+loss:
+  chamfer_weight: 1.0
+  tracked_l2_weight: 1.0
+
+material_guess:
+  mode: uniform
+  log_E_range: [7.66, 7.66]
+  nu_range: [0.363, 0.363]
+  log_E_clamp_range: [5.0, 12.0]
+  nu_clamp_range: [0.3, 0.43]

configs/validate_MfM.yaml

@@ -0,0 +1,150 @@
+# MfM (Material from Motion) validation. Owned by validate_MfM.py.
+# `refiner` and `Refiner`, here and in the code, both mean MfM.
+
+validation:
+  checkpoint: checkpoints/MfM_checkpoint.pt
+  device: cuda
+  # Report path, stamped by hand at launch. Saved trajectories are named after
+  # it, so null turns trajectory saving off even when save_rollout is true.
+  out: outputs/MfM/val/YYYYMMDD_hhmm/validation.json
+  validation_mode: pure-tail   # `full` and `refresh-tail` are ablations
+  save_rollout: true
+
+dataset:
+  validation_roots:
+    - data_episodes/physcore/*/episode_0000
+  rollout_steps: 25
+  cache_size: 1
+  real_world_domain_center: [0.5, 0.5, 0.2]
+  observation_views: all
+  observation_view_index: 0
+
+# Refiner (MfM) architecture. Must match the checkpoint being validated, since the
+# model is rebuilt from this block before the weights load.
+model:
+  hidden: 128
+  graph_unet_k: 8
+  graph_unet_pool: fps
+  graph_unet_ratios:
+  - 1.0
+  - 0.5
+  - 0.25
+  edge_feature: canonical_length_only
+  edge_direction: false
+  fixed_tracked_mask: false
+  use_persistent_tracks: true
+  persistent_track_use_motion_valid: true
+  tracked_disp_mode: canonical
+  control_disp_mode: absolute
+  max_controls: 4
+  tracked_disp_scale: 33.0
+  tracked_mask_scale: 1.0
+  control_vec_scale: 1.0
+  control_disp_scale: 33.0
+  canonical_bands: 4
+  tracked_bands: 16
+  attention_window: 9
+  temporal_conv: 9
+  initial_material_confidence_bias: 0.0
+  evidence_aggregation: true
+  evidence_tracked_weight: 8.0
+  evidence_control_weight: 4.0
+  evidence_control_radius: 0.04
+  log_E_min: 5.0
+  log_E_max: 11.0
+  nu_min: 0.05
+  nu_max: 0.45
+  graph_aggregation_chunks: 10
+
+# The only two training fields validation reads.
+train:
+  seed: 0
+  update_every: 10
+
+# Initial material the rollout starts from; the clamps bound the predicted values.
+material_guess:
+  mode: uniform
+  log_E_range: [7.66, 7.66]
+  nu_range: [0.363, 0.363]
+  log_E_clamp_range: [5.0, 12.0]
+  nu_clamp_range: [0.05, 0.45]
+
+rollout:
+  use_batched_mpm: true
+  use_episode_sim_dt: true
+  default_dt: 0.000666
+  default_steps_per_frame: 2
+  num_grids: 32
+  clip_bound: 0.5
+  gravity: [0.0, 0.0, -9.8]
+  damping: 0.999
+  rho: 100.0
+  ground_height: 0.02
+  ground_surface: coulomb
+  ground_elasticity: 0.0
+  ground_friction: 0.5
+  ground_tangent_damping: 0.25
+  ground_static_velocity_threshold: 0.02
+  use_explicit_rigid_contact: true
+  use_episode_rigid_contact: true
+  rigid_surface_fps: 256
+  rigid_friction: 0.5
+  rigid_contact_margin: 0.02
+  rigid_velocity_blend: 0.9
+  rigid_stickiness: 1.0
+  rigid_strict_nonpenetration: true
+  rigid_projection_iterations: 2
+  rigid_projection_margin: 5.0e-05
+  elasticity_model: CorotatedElasticity
+  # Elastic by default; plastic cases override it in per_sample_rollout_config.
+  plasticity_model: IdentityPlasticity
+  manipulation_controller_grid_contact_radius: 0.04  # tune per case (scene-dependent)
+  manipulation_controller_grid_velocity_blend: 1.0
+  max_delta_velocity_norm: 0.05
+  correction_target_velocity_gain: 100.0
+  correction_target_velocity_damping: 1.2
+  correction_target_velocity_integral_gain: 40.0
+  correction:
+    mode: spring
+    dt_scale: 0.25
+    steps_per_frame_scale: 4
+    velocity_cap: 5.0
+    omega_n: 150.0
+    zeta: 2.5
+  apply_force_every_substep: true
+  max_force_norm: 5000.0
+
+per_sample_rollout_config:
+  double_squeeze_plastic:
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+    manipulation_controller_grid_contact_radius: 0.04
+  double_stretch_bear:
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+    manipulation_controller_grid_contact_radius: 0.04
+  double_stretch_bear_1:
+    plasticity_model: IdentityPlasticity
+    manipulation_controller_grid_contact_radius: 0.06
+  double_stretch_plastic:
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+    manipulation_controller_grid_contact_radius: 0.04
+  single_lift_rope:
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+    manipulation_controller_grid_contact_radius: 0.03
+  single_push_rope:
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+    manipulation_controller_grid_contact_radius: 0.02

configs/validate_RfD.yaml

@@ -0,0 +1,205 @@
+# RfD (Residual from Dynamics) validation sweep. Owned by validate_RfD.py.
+# `gvc` = legacy RfD, `refiner` = MfM. train.output_dir must match train_RfD.yaml;
+# every checkpoint there is scored on the held-out second half of each episode.
+
+dataset:
+  train_roots:
+    - data_episodes/physcore/double_clift_cloth/episode_0000
+    - data_episodes/physcore/double_fold_rope/episode_0000
+    - data_episodes/physcore/double_squeeze_plastic/episode_0000
+    - data_episodes/physcore/double_stretch_bear/episode_0000
+    - data_episodes/physcore/double_stretch_bear_1/episode_0000
+    - data_episodes/physcore/double_stretch_plastic/episode_0000
+    - data_episodes/physcore/single_clift_cloth/episode_0000
+    - data_episodes/physcore/single_clift_rope/episode_0000
+    - data_episodes/physcore/single_lift_bear/episode_0000
+    - data_episodes/physcore/single_lift_cloth/episode_0000
+    - data_episodes/physcore/single_lift_rope/episode_0000
+    - data_episodes/physcore/single_push_rope/episode_0000
+  validation_roots:
+    - data_episodes/physcore/double_clift_cloth/episode_0000
+    - data_episodes/physcore/double_fold_rope/episode_0000
+    - data_episodes/physcore/double_squeeze_plastic/episode_0000
+    - data_episodes/physcore/double_stretch_bear/episode_0000
+    - data_episodes/physcore/double_stretch_bear_1/episode_0000
+    - data_episodes/physcore/double_stretch_plastic/episode_0000
+    - data_episodes/physcore/single_clift_cloth/episode_0000
+    - data_episodes/physcore/single_clift_rope/episode_0000
+    - data_episodes/physcore/single_lift_bear/episode_0000
+    - data_episodes/physcore/single_lift_cloth/episode_0000
+    - data_episodes/physcore/single_lift_rope/episode_0000
+    - data_episodes/physcore/single_push_rope/episode_0000
+  rollout_steps: 25
+  cache_size: 1
+  real_world_domain_center: [0.5, 0.5, 0.2]
+  observation_views: all
+  observation_view_index: 0
+
+# MPM rollout physics. Identical to configs/train_RfD.yaml, base radius and
+# per-case overrides alike, so each case is scored as it was adapted.
+rollout:
+  use_batched_mpm: true
+  use_episode_sim_dt: true
+  default_dt: 0.000666
+  default_steps_per_frame: 2
+  num_grids: 32
+  clip_bound: 0.5
+  gravity: [0.0, 0.0, -9.8]
+  damping: 0.999
+  rho: 100.0
+  ground_height: 0.02
+  ground_surface: coulomb
+  ground_elasticity: 0.0
+  ground_friction: 0.5
+  ground_tangent_damping: 0.25
+  ground_static_velocity_threshold: 0.02
+  use_explicit_rigid_contact: true
+  use_episode_rigid_contact: true
+  rigid_surface_fps: 256
+  rigid_friction: 0.5
+  rigid_contact_margin: 0.02
+  rigid_velocity_blend: 0.9
+  rigid_stickiness: 1.0
+  rigid_strict_nonpenetration: true
+  rigid_projection_iterations: 2
+  rigid_projection_margin: 5.0e-05
+  elasticity_model: CorotatedElasticity
+  plasticity_model: IdentityPlasticity
+  manipulation_controller_grid_contact_radius: 0.04
+  manipulation_controller_grid_velocity_blend: 1.0
+  max_delta_velocity_norm: 0.05
+  correction_target_velocity_gain: 100.0
+  correction_target_velocity_damping: 1.2
+  correction_target_velocity_integral_gain: 40.0
+  correction:
+    mode: spring
+    dt_scale: 0.25
+    steps_per_frame_scale: 4
+    velocity_cap: 5.0
+    omega_n: 150.0
+    zeta: 2.5
+  apply_force_every_substep: true
+  max_force_norm: 5000.0
+
+# Per-case rollout overrides, keyed by the case (parent directory) name. These
+# match configs/train_RfD.yaml, so each case is scored as it was adapted.
+per_sample_rollout_config:
+  double_squeeze_plastic:
+    manipulation_controller_grid_contact_radius: 0.08  # tune per case (scene-dependent)
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  double_stretch_bear:
+    manipulation_controller_grid_contact_radius: 0.08  # matches the submitted run
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  double_stretch_bear_1:
+    # Radius only, so this case keeps the base IdentityPlasticity.
+    manipulation_controller_grid_contact_radius: 0.06
+  double_stretch_plastic:
+    manipulation_controller_grid_contact_radius: 0.08  # tune per case (scene-dependent)
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  single_lift_rope:
+    manipulation_controller_grid_contact_radius: 0.03  # tune per case (scene-dependent)
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+  single_push_rope:
+    manipulation_controller_grid_contact_radius: 0.02  # tune per case (scene-dependent)
+    plasticity_model: VonMisesPlasticity
+    plasticity_kwargs:
+      VonMisesPlasticity:
+        sigma_y: 200.0
+
+# Refiner (MfM) architecture. The checkpoint carries no embedded cfg, so the
+# model is built from this block and loaded strictly. Keep it in sync with
+# configs/train_RfD.yaml and gvc.refiner_checkpoint.
+model:
+  spatial: graph_unet
+  temporal: conv_gru_fusion
+  hidden: 128
+  layers: 4
+  k: 8
+  graph_unet_k: 8
+  graph_unet_pool: fps
+  graph_unet_ratios: [1.0, 0.5, 0.25]
+  edge_feature: canonical_length_only
+  edge_direction: false
+  input_mode: observed_control
+  fixed_tracked_mask: false
+  use_persistent_tracks: true
+  persistent_track_use_motion_valid: true
+  tracked_disp_mode: canonical
+  control_disp_mode: absolute
+  max_controls: 4
+  observed_drop_canonical: false
+  tracked_disp_scale: 33.0
+  tracked_mask_scale: 1.0
+  control_vec_scale: 1.0
+  control_disp_scale: 33.0
+  coord_bands: 16
+  canonical_bands: 4
+  tracked_bands: 16
+  control_bands: 1
+  attention_window: 9
+  attention_heads: 4
+  temporal_conv: 9
+  use_material: false
+  use_correction: false
+  use_correction_mask: false
+  use_velocity: false
+  use_F: false
+  use_material_confidence: true
+  separate_confidence: true
+  detached_confidence: false
+  initial_material_confidence_bias: 0.0
+  aggregate_controls: true
+  evidence_aggregation: true
+  evidence_tracked_weight: 8.0
+  evidence_control_weight: 4.0
+  evidence_control_radius: 0.04
+  log_E_min: 5.0
+  log_E_max: 11.0
+  nu_min: 0.05
+  nu_max: 0.45
+  graph_aggregation_chunks: 10
+  use_decomposed_material: false
+
+# RfD, still spelled gvc here.
+gvc:
+  base_channels: 32
+  levels: 2
+  max_delta_v: 0.05        
+  h: 10                    # corrector fires every h MPM substeps
+  warmup_fraction: 0.5     # first-half boundary, shared with configs/train_RfD.yaml
+  refiner_checkpoint: checkpoints/MfM_checkpoint.pt
+
+train:
+  seed: 42
+  device: cuda
+  # Run directory of the training run being validated. Must equal
+  # train.output_dir in configs/train_RfD.yaml, or the sweep aborts with
+  # "checkpoints dir not found".
+  output_dir: outputs/RfD/YYYYMMDD_hhmm
+  update_every: 10         # k = window size in camera frames
+  viz_every: 20
+  viz_fps: 30
+  viz_max_controls: 16
+
+loss:
+  chamfer_weight: 1.0
+  tracked_l2_weight: 1.0
+
+material_guess:
+  mode: uniform
+  log_E_range: [7.66, 7.66]
+  nu_range: [0.363, 0.363]
+  log_E_clamp_range: [5.0, 12.0]
+  nu_clamp_range: [0.3, 0.43]

data/physcore.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e6669c77c77e0e68ddc6f0fd90704eba3bc35d2af8b1db885063015b91bc40c
+size 4208002608

data/phystwin.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f658b0a4f520721e66da38ea070c73ce5b43b2f22bbc18d8931ee1f96a5a019
+size 3441847915

gaussian_output/physcore.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c498c12fd09848063e09fa0a792afe99899b4d8833dcdf30be2671d80808a00
+size 111477552