CIRL Demonstration Dataset

This repository provides demonstration datasets for 12 robotic manipulation tasks used in CIRL (Causal Reinforcement Learning). Each task contains 10 expert demonstration trajectories stored as sequential transitions with synchronized multi-view vision (external camera + wrist camera) and robot proprioceptive states.

---

📊 Dataset Summary

• Total Tasks: 12 diverse manipulation tasks
• Demonstrations per Task: 10 trajectories
• Total Transitions: 7,404 step transitions
• Total Size: ~1.46 GB
• Storage Format: Python pickle (.pkl) serialized transition lists

---

🗂️ Dataset Structure & Tasks

All dataset files are located in the CIRL_demodata/ folder:

File Name	Task Description	Demos	Transitions	File Size
assemble_ram_10_demos_*.pkl	Align and insert RAM card into slot	10	540	~101 MB
fold_towel_10_demos_*.pkl	Deformable object folding (towel)	10	907	~170 MB
insert_cube_10_demos_*.pkl	Precision insertion of cube into container	10	273	~51 MB
insert_plug_10_demos_*.pkl	Grasp and insert plug into socket	10	375	~70 MB
open_toaster_10_demos_*.pkl	Articulated object interaction (toaster)	10	495	~93 MB
pick_among_items_10_demos_*.pkl	Target selection and picking in clutter	10	349	~65 MB
pick_eggplant_10_demos_*.pkl	Pick target object (eggplant)	10	339	~64 MB
pick_place_banana_10_demos_*.pkl	Pick and relocate banana to target	10	727	~137 MB
put_toast_10_demos_*.pkl	Pick and place toast into appliance	10	1,636	~307 MB
stack_block_10_demos_*.pkl	Multi-stage stacking of blocks	10	864	~162 MB
touch_eggplant_10_demos_*.pkl	Reaching / touching target object	10	287	~54 MB
wipe_whiteboard_10_demos_*.pkl	Surface wiping motion along trajectory	10	611	~115 MB

---

📐 Data Schema

Each .pkl file contains a Python list of transition dictionaries. Each transition represents a single timestep $(s_t, a_t, s_{t+1}, r_t, d_t)$ formatted as follows:

{
    "observations": {
        "external": np.ndarray,      # shape: (1, 128, 128, 3), dtype: uint8 (Third-person view)
        "wrist":    np.ndarray,      # shape: (1, 128, 128, 3), dtype: uint8 (Eye-in-hand view)
        "state":    np.ndarray,      # shape: (1, 11),          dtype: float32 (Proprioceptive state)
    },
    "actions": np.ndarray,           # shape: (3,),             dtype: float64 (Continuous control)
    "next_observations": {
        "external": np.ndarray,      # shape: (1, 128, 128, 3), dtype: uint8
        "wrist":    np.ndarray,      # shape: (1, 128, 128, 3), dtype: uint8
        "state":    np.ndarray,      # shape: (1, 11),          dtype: float32
    },
    "rewards": float,                # Task reward at current step
    "masks": float,                  # 1.0 if continuing, 0.0 if episode ends
    "dones": bool,                   # Episode termination indicator
    "infos": {
        "executed_action": np.ndarray, # Executed action array
        "left": bool,
        "right": bool
    }
}