Update README for 2026-04-23 recollection

README.md

@@ -1,151 +1,132 @@
 ---
 license: mit
 task_categories:
-  - robotics
-tags:
-  - physics
-  - dynamics
-  - probing
-  - isaac-lab
-  - manipulation
-  - physrepa
-  - benchmark
-pretty_name: Dynamics Probing Benchmark Dataset
+- robotics
+- reinforcement-learning
+language:
+- en
+pretty_name: PhysProbe Dynamics Probing Dataset
 size_categories:
-  - 10K<n<100K
+- 10K<n<100K
+tags:
+- robotics
+- isaac-lab
+- physics
+- probing
+- manipulation
+- franka
 ---
 
-# Dynamics Probing Benchmark Dataset
-
-Physics-varied manipulation episodes for probing whether video encoders encode contact dynamics in their intermediate representations.
-
-## Overview
+# PhysProbe Dynamics Probing Dataset
 
-This dataset contains **12,100 episodes** across 6 manipulation tasks in Isaac Lab simulation, with **per-episode physics randomization** (friction, mass, damping, restitution). Each episode includes dual-camera RGB video (384x384), robot state/action trajectories, and dense physics ground truth annotations.
-
-**Purpose**: Dynamics probing — train linear probes on frozen video encoder features to predict physics quantities (contact force, friction coefficient, object velocity, etc.) and identify which layers encode contact dynamics.
+Manipulation episodes from Isaac Lab collected for probing physics understanding in video world models (V-JEPA 2, VideoMAE, DINOv2). Each episode includes dual-camera RGB (384×384), robot state, scripted/RL actions, and per-timestep physics ground truth (contact forces, object kinematics, physics randomization parameters).
 
 ## Tasks
 
-| Task | Episodes | Policy | Physics Params | Description |
-|------|----------|--------|----------------|-------------|
-| **Push** | 1,500 | Scripted (Step0) | mass, obj friction, surface friction | Franka pushes cube on table, random direction |
-| **Strike** | 3,000 | Scripted (Step0) | mass, obj friction, surface friction, damping, restitution | Franka strikes ball, random direction |
-| **PegInsert** | 2,500 | RL (rl_games) | held friction, fixed friction, held mass | Peg-in-hole insertion with pre-grasped peg |
-| **NutThread** | 2,500 | RL (rl_games) | held friction, fixed friction, held mass | Nut threading onto bolt with pre-grasped nut |
-| **Drawer** | 2,000 | RL (RSL-RL) | handle friction, gripper friction, drawer damping | Open cabinet drawer |
-| **Reach** | 600 | Scripted (Oracle) | None (negative control) | Move EE to random target position |
+| Task | Episodes | Policy | Physics Randomization |
+|------|---------:|--------|------------------------|
+| Push | 1,500 | Scripted (random direction, no target — Step 0) | mass, obj_friction, surface_friction |
+| Strike | 3,000 | Scripted (random direction, no target — Step 0) | mass, friction, surface_friction, restitution |
+| Reach | 600 | Scripted | None (negative control) |
+| Drawer | 2,000 | RL (RSL_RL) | drawer_joint_damping |
+| PegInsert | 2,500 | Scripted (Factory) | held_friction, fixed_friction, held_mass |
+| NutThread | 2,500 | Scripted (Factory) | held_friction, fixed_friction, held_mass |
+| **Total** | **12,100** | | |
 
-## Data Format (LeRobot V2)
+## Format
 
-Each task directory contains:
+LeRobot V2 layout. Per-task:
 ```
-{task}/
-├── data/chunk-000/
-│   ├── episode_000000.parquet
-│   ├── episode_000001.parquet
-│   └── ...
-├── videos/chunk-000/
-│   ├── observation.images.image_0/  (table camera)
-│   └── observation.images.image_1/  (wrist camera)
-└── meta/
-    ├── info.json
-    ├── episodes.jsonl
-    ├── tasks.jsonl
-    └── stats.json
+<task>/
+  data/chunk-000/episode_NNNNNN.parquet
+  videos/chunk-000/observation.images.image_0/episode_NNNNNN.mp4   # table_cam
+  videos/chunk-000/observation.images.image_1/episode_NNNNNN.mp4   # wrist_cam
+  meta/info.json
+  meta/episodes.jsonl
+  meta/tasks.jsonl
+  meta/stats.json
+  meta/modality.json
 ```
 
-### Per-frame columns (parquet)
-
-| Column | Shape | Description |
-|--------|-------|-------------|
-| `observation.state` | (8,) | EE pose [x,y,z,r,p,y,pad,gripper] |
-| `action` | (7,) | IK delta [dx,dy,dz,dr,dp,dy,gripper] |
-| `observation.images.image_0` | video | Table camera RGB 384x384 |
-| `observation.images.image_1` | video | Wrist camera RGB 384x384 |
+Per-episode parquet columns:
+- `observation.state` (8D): 7 joint positions + 1 gripper
+- `action` (3–8D, task-dependent)
+- `next.reward`, `next.done`
+- `physics_gt.*` (task-specific — see below)
+- Frame index, timestep, episode index metadata
 
-### Physics Ground Truth (per-frame, task-specific)
+## Physics Ground Truth (`physics_gt.*`)
 
-**Common (all tasks)**:
-- `physics_gt.ee_position` (3), `ee_orientation` (4), `ee_velocity` (3), `ee_angular_velocity` (3), `ee_acceleration` (3)
+### Common across all tasks
+- `ee_position` (3), `ee_orientation` (4), `ee_velocity` (3), `ee_angular_velocity` (3) — end-effector kinematics
+- `joint_pos` (7), `joint_vel` (7) — arm joint state
+- `phase` (1) — task phase label (task-dependent enum; 7 = idle)
 
-**Push/Strike**:
-- `physics_gt.object_position` (3), `object_velocity` (3), `object_angular_velocity` (3), `object_acceleration` (3)
-- `physics_gt.contact_finger_l_object_flag` (1), `contact_finger_l_object_force` (3)
-- `physics_gt.contact_object_surface_flag` (1), `contact_object_surface_force` (3)
-- `physics_gt.object_to_target_distance` (1), `object_on_surface` (1)
-- Strike only: `physics_gt.ball_planar_travel_distance` (1)
+### Contact fields (per task)
 
-**PegInsert**:
-- `physics_gt.peg_position` (3), `peg_orientation` (4), `peg_velocity` (3), `peg_angular_velocity` (3)
-- `physics_gt.hole_position` (3)
-- `physics_gt.insertion_depth` (1), `peg_hole_lateral_error` (1)
-- Pair contacts: `contact_finger_l_peg_*`, `contact_finger_r_peg_*`, `contact_peg_socket_*`
+**Push, Strike, Reach:**
+- `contact_flag` (1), `contact_force` (3), `contact_point` (3) — aggregate ee↔object + object↔surface
+- `contact_finger_l_object_flag/force`, `contact_finger_r_object_flag/force` — per-finger ee↔object (new in 2026-04-23)
+- `contact_object_surface_flag/force` — object↔surface (new in 2026-04-23)
 
-**NutThread**:
-- `physics_gt.nut_position` (3), `nut_orientation` (4), `nut_velocity` (3), `nut_angular_velocity` (3)
-- `physics_gt.bolt_position` (3), `bolt_orientation` (4)
-- `physics_gt.axial_progress` (1), `nut_bolt_relative_angle` (1)
-- Pair contacts: `contact_finger_l_nut_*`, `contact_finger_r_nut_*`, `contact_nut_bolt_*`
+**PegInsert, NutThread:**
+- `contact_flag` (1), `contact_force` (3), `contact_point` (3) — aggregate
+- `contact_finger_l_peg_flag/force`, `contact_finger_r_peg_flag/force` (PegInsert) — finger↔peg
+- `contact_finger_l_nut_flag/force`, `contact_finger_r_nut_flag/force` (NutThread) — finger↔nut
+- `contact_peg_socket_flag/force` (PegInsert) — peg↔hole (reconstructed as residual from peg total contact minus finger reactions; direct pair filter unsupported in Factory direct env)
+- `contact_nut_bolt_flag/force` (NutThread) — nut↔bolt (direct exact-pair sensor)
 
-**Drawer**:
-- `physics_gt.drawer_joint_pos` (1), `drawer_joint_vel` (1)
-- `physics_gt.handle_position` (3), `handle_velocity` (3)
-- `physics_gt.drawer_opening_extent` (1)
-- Pair contacts: `contact_finger_l_handle_*`, `contact_finger_r_handle_*`
+**Drawer:**
+- `handle_position` (3), `handle_velocity` (3)
+- `drawer_joint_pos` (1), `drawer_joint_vel` (1)
 
-**Reach** (negative control):
-- `physics_gt.target_position` (3), `ee_to_target_distance` (1)
+### Task-specific kinematics
 
-### Episode Metadata (episodes.jsonl)
+**Push/Strike/Reach (Step 0):**
+- `object_position` (3), `object_orientation` (4), `object_velocity` (3), `object_angular_velocity` (3)
+- `target_position` (3) — placeholder `[0, 0, 0]` (no target in Step 0)
 
-Each episode records randomized physics parameters:
-- Push/Strike: `object_0_mass`, `object_0_static_friction`, `surface_static_friction`
-- PegInsert: `peg_static_friction`, `peg_mass`, `hole_static_friction`
-- NutThread: `nut_static_friction`, `nut_mass`, `bolt_static_friction`
-- Drawer: `drawer_joint_damping`, `handle_static_friction`
+**PegInsert/NutThread:**
+- `held_position` (3), `held_orientation` (4), `held_velocity` (3), `held_angular_velocity` (3) — peg/nut kinematics
+- `fixed_position` (3), `fixed_orientation` (4) — hole/bolt pose
+- `insertion_depth` (1) — peg_insert only
 
-## Environment
+### Physics randomization parameters (per episode)
 
-- **Simulator**: Isaac Lab v2.2.1 on Isaac Sim 4.5.0
-- **Robot**: Franka Panda (all tasks)
-- **Control**: Differential IK (position-only for Push/Strike, pose for others)
-- **Cameras**: 384x384 RGB, table_cam (third-person) + wrist_cam
-- **Physics**: PhysX GPU, dt=0.01s, decimation=2 (50Hz control)
+Stored in episode metadata (`physics_gt.*_{static,dynamic}_friction`, `*_mass`, etc.) — see per-task schema above for exact fields.
 
-## Design Rationale
+## 2026-04-23 Recollection Note
 
-**Why physics randomization?** If a video encoder encodes physics in its intermediate layers, linear probes trained on episodes with varied friction/mass should predict these quantities from frozen features. Tasks without physics variation (Reach) serve as negative controls.
+The previous version of this dataset (before 2026-04-23) had a data-collection bug: contact forces were zero-filled across all tasks because the sensor configuration did not use the proper body filters / the Factory direct env does not support `get_net_contact_forces` on `ArticulationView`. This version fixes the following:
 
-**Episode count rationale**: Proportional to `(param space dim) × (dynamics complexity) / (informative frame density)`. Strike has 3,000 episodes (most) due to 5D param space and sparse impact events (~25% informative frames).
+1. **Push, Strike, Drawer, Reach**: Per-pair `ContactSensor` with `filter_prim_paths_expr` on finger/object bodies → real nonzero contact forces.
+2. **NutThread**: Direct exact-pair sensor (`contact_nut_bolt_*`) → direct nut↔bolt force.
+3. **PegInsert**: GPU pair filtering on hole is unsupported in direct Factory env. Peg↔socket contact is reconstructed as a **residual**: `F_peg_socket = F_peg_total - F_finger_l_peg - F_finger_r_peg` (Newton's 3rd law). This is sparser and noisier than a direct sensor; finger-grip force dominates and is subtracted, so pay attention when using `contact_peg_socket_*` for downstream probing.
+4. **Phase label**: Now correctly tracks scripted policy state transitions (previously always 7/idle for RL policies).
 
-## Reproducibility
+### Known caveats (unchanged from previous release)
 
-- **Code**: [PhysREPA_Tasks](https://github.com/Leesangoh/PhysREPA_Tasks) (commit `2808943`)
-- **RL Checkpoints**: Included in `checkpoints/` directory of this repo
-  - `peg_insert_rl_games.pth` — PegInsert (rl_games, reward 381)
-  - `nut_thread_rl_games.pth` — NutThread (rl_games, reward 952)
-  - `drawer_rsl_rl.pt` — Drawer (RSL-RL, model_4999)
-- **Environment**: `environment_info.json` and `pip_freeze.txt` in this repo
-- **Platform**: Isaac Sim 4.5.0, Isaac Lab v2.2.1, Python 3.10, CUDA 12.8, 4x NVIDIA A6000
+- Push/Strike are Step 0: no target, `success=True` for all, `target_position=[0,0,0]`.
+- Drawer randomizes only `drawer_joint_damping` (Isaac Lab env limitation — handle friction/mass are fixed).
+- Factory tasks (PegInsert, NutThread): collection uses scripted policy; success rate is low for unguided inserts.
 
-### Reproduce collection
+## Collection Pipeline
 
-```bash
-# Clone code
-git clone https://github.com/Leesangoh/PhysREPA_Tasks
-cd PhysREPA_Tasks
+- Isaac Sim 4.5 / Isaac Lab v2.2.1
+- Scripted oracle policies + optional RL checkpoints (`rl_games` for Factory, `rsl_rl` for Drawer)
+- Dual camera rendering at 384×384, 15 fps
+- `num_envs` per task: 8 (Factory), 16 (others) parallel
+- Hardware: 4× NVIDIA A6000 48 GB
 
-# Push (scripted, 16 parallel envs)
-/isaac-sim/python.sh collect_sample_data.py --task push --num_episodes 1500 --num_envs 16 --step0 --output_dir <OUT> --headless
+Collection script: [Leesangoh/PhysREPA_Tasks](https://github.com/Leesangoh/PhysREPA_Tasks) (`archive_data_collection/collect_sample_data.py`).
 
-# PegInsert (RL, 200 parallel envs)
-/isaac-sim/python.sh collect_sample_data.py --task peg_insert --num_episodes 2500 --num_envs 200 --filter_success --rl_checkpoint checkpoints/peg_insert_rl_games.pth --output_dir <OUT> --headless
+## Intended use
 
-# Drawer (RL, 32 parallel envs)
-/isaac-sim/python.sh collect_sample_data.py --task drawer --num_episodes 2000 --num_envs 32 --filter_success --rl_checkpoint checkpoints/drawer_rsl_rl.pt --output_dir <OUT> --headless
-```
+This dataset is designed for **probing physics understanding** in pretrained video encoders:
+- Linear probes from mean-pooled features onto `physics_gt.*` targets
+- Temporal-aligned window-level probing (per-window features → per-window targets)
+- Do NOT episode-mean aggregate features/targets for kinematic probing — that collapses temporal structure and produces misleading results.
 
-## License
+## Citation
 
-MIT
+TBD (paper in preparation).