--- title: "Zero-Trust Physics: 1000Hz Domain-Randomized Humanoid Grasp & Slip Trajectories (ZTP-TSA-v1)" dataset_info: features: - name: episode_id dtype: int64 - name: timestamp_sec dtype: float64 - name: mass_kg dtype: float64 - name: true_static_mu dtype: float64 - name: actual_normal_force_n dtype: float64 - name: slip_velocity_mps dtype: float64 - name: slip_angular_velocity_radps dtype: float64 - name: slip_displacement_mm dtype: float64 - name: rot_slip_displacement_deg dtype: float64 - name: action_commanded_force_n dtype: float64 - name: estimated_mu dtype: float64 - name: safety_margin dtype: float64 - name: micro_slip_detected dtype: bool - name: macro_slip_detected dtype: bool - name: rotational_slip_detected dtype: bool - name: taxel_0_normal_n dtype: float64 # Note: Includes taxel_0 to taxel_15 normal, shear_x, and shear_y forces (64 features total) splits: - name: train num_bytes: 224000000 num_examples: 438605 download_size: 12270000 dataset_size: 224000000 configs: - config_name: default data_files: - split: train path: haptic_trajectories.parquet tags: - robotics - reinforcement-learning - tactile-sensing - haptic-feedback - humanoid-manipulation - zero-trust-physics - aegis-os - domain-randomization --- # 🤖 Zero-Trust Physics: 1000Hz Dexterous Grasp & Tactile Slip Trajectories (ZTP-TSA-v1) [![Domain](https://img.shields.io/badge/Website-ZeroTrustPhysics.com-blue?style=flat-square)](https://ZeroTrustPhysics.com) [![Maintainer](https://img.shields.io/badge/Lead--Architect-Kruze-purple?style=flat-square)](mailto:kruze@zerotrustphysics.com) This dataset contains **high-fidelity 1000Hz timeseries telemetry** representing dynamic humanoid robotic grasps under translational and torsional slip disturbances. The data was generated by the **Zero-Trust Physics Tactile Slip Auditor (ZTP-TSA)**, a zero-dependency, FPU-optimized, `no-std` Rust solver. ## ⚡ Why This Dataset is Different Standard robotic datasets capture joint states or RGB images at 30-60Hz. **This dataset captures tactile pressure and shear distribution vectors at 1000Hz (1.0ms resolution)**. It exposes the physical transition from **micro-slip** (local tangential sliding of boundary contact points) to **macro-slip** (global translation sliding) and **rotational slip** (circumferential slipping under torsional torque), alongside real-time reflex force corrections executed by an edge microcontroller solver running at FPU speeds. ``` [4x4 Tactile Sensor Array] (Normal & Shear x/y at 1000Hz) │ ▼ [no-std Rust Solver: ZTP-TSA] (Runs in < 10 microseconds) │ ┌────────────┴────────────┐ ▼ ▼ [Friction Estimation] [Reflex Correction] (Adapts mu_s in 5ms) (Arrests slip in < 2ms) ``` --- ## 📊 Dataset Schema & Columns Each row in the dataset corresponds to a single **1.0 millisecond control cycle** ($dt = 0.001\text{ s}$). The schema is structured for **offline Reinforcement Learning (RL), Behavior Cloning (BC), or VLA (Vision-Language-Action) haptic policy pretraining**. ### 1. Episode / Metadata * `timestamp_sec` (float64): Timestamp of the control cycle since episode start. * `scenario` (string): The physical environment parameters: * `NOMINAL`: $0.8\text{ kg}$ object, dry tactile pad ($\mu_s = 0.60$). * `HEAVY_LOAD`: $1.0\text{ kg}$ object, high acceleration load ($\mu_s = 0.60$). * `OILED_SURFACE`: $0.4\text{ kg}$ object, sudden low-friction oil patch ($\mu_s = 0.25 \rightarrow 0.10$). ### 2. Ground-Truth Physics State (`state`) * `mass_kg` (float64): Mass of the grasped object. * `actual_normal_force_n` (float64): Actual normal force exerted by the actuator finger pad. * `slip_velocity_mps` (float64): Linear slip velocity relative to the finger pad (meters/sec). * `slip_angular_velocity_radps` (float64): Rotational slip velocity (radians/sec). * `slip_displacement_mm` (float64): Cumulative linear slip distance (millimeters). * `rot_slip_displacement_deg` (float64): Cumulative rotational slip angle (degrees). * `static_friction_coeff` / `dynamic_friction_coeff` (float64): True physical friction coefficients. ### 3. High-Density Tactile Sensor Array (`sensor`) A coordinate-mapped $4 \times 4$ tactile matrix centered at index $(1.5, 1.5)$. For each of the **16 taxels** ($i \in [0, 15]$): * `taxel_{i}_normal_n` (float64): Perpendicular normal pressure force. * `taxel_{i}_shear_x_n` (float64): Tangential shear force along the contact grid horizontal axis. * `taxel_{i}_shear_y_n` (float64): Tangential shear force along the contact grid vertical axis. ### 4. Reflex Controller Action (`reflex`) * `action_commanded_force_n` (float64): Commanded target normal force output by the ZTP-TSA controller. * `estimated_mu` (float64): Real-time friction coefficient ($\mu_s$) computed by the running observer. * `safety_margin` (float64): Friction cone safety margin ($1.0 = \text{no shear}$, $0.0 = \text{imminent slip}$). * `micro_slip_detected` (bool): Flagged when boundary contact taxels exceed the local friction limit. * `macro_slip_detected` (bool): Flagged when global linear slip is detected. * `rotational_slip_detected` (bool): Flagged when torsional moment exceeds the contact boundary limit. --- ## 🏎️ Physical Benchmarks & Catch Latencies The underlying Rust FFI engine stabilizes all three scenarios in **under 2.0ms** of slip onset. | Scenario | Object Mass | Start Force | Friction Envelope | Catch Latency | Final Slip | Status | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | **NOMINAL** | $0.8\text{ kg}$ | $10.0\text{ N}$ | $\mu_s = 0.60$ | **1.00 ms** | $3.85\text{ mm}$ | **Secured** | | **HEAVY LOAD** | $1.0\text{ kg}$ | $12.0\text{ N}$ | $\mu_s = 0.60$ | **1.00 ms** | $15.39\text{ mm}$ | **Secured** | | **OILED SURFACE**| $0.4\text{ kg}$ | $14.0\text{ N}$ | $\mu_s = 0.25 \rightarrow 0.10$ | **1.00 ms** | $9.48\text{ mm}$ | **Secured** | *Note: Latency calculations include a realistic $1.5\text{ ms}$ physical actuator lag ($\tau = 0.0015\text{ s}$) matching direct-drive humanoid motor dynamics.* --- ## 🛠️ Loading the Dataset ### Python (using Hugging Face `datasets`) ```python from datasets import load_dataset import pandas as pd # Stream or download the haptic dataset dataset = load_dataset("spiderpilot89/humanoid-tactile-slip-reflex-1000hz", split="train") # Convert to Pandas DataFrame df = dataset.to_pandas() # Filter for the oiled surface patch scenario oiled_data = df[df["scenario"] == "OILED_SURFACE"] # View tactile array shear vector for the first taxel print(oiled_data[["timestamp_sec", "taxel_0_normal_n", "taxel_0_shear_x_n", "taxel_0_shear_y_n"]].head()) ``` --- ## 🔬 Mathematical Formulations ### 1. Coordinate-Mapped Torsional Moment The solver maps each taxel index $i$ to a 2D physical spatial vector $(dx_i, dy_i)$ relative to the array center. The net torsional moment $M_z$ is computed as: $$M_z = \sum_{i=0}^{15} \left( dx_i \cdot \text{shear\_y}_i - dy_i \cdot \text{shear\_x}_i \right)$$ ### 2. Adaptive Friction Observer When boundary taxels begin to slip, their tangential-to-normal force ratio is calculated. The system adapts its friction limits dynamically using a first-order low-pass filter ($\alpha = 0.05$): $$\mu_{s, \text{new}} = \mu_{s, \text{old}} \cdot (1 - \alpha) + \mu_{\text{measured}, \text{avg}} \cdot \alpha$$ --- ## 🔒 Cryptographic Audit & Aegis OS Seal Every simulation run generates a cryptographic proof sealed by **Aegis OS**. The integrity hash protects the simulation manifest from tampering, ensuring that the control latency parameters presented match real execution cycles: * **Domain ID**: `humanoid_dexterity` * **Sovereignty Seal (SHA256)**: `ea0fd0d7c71d6f1f4405a3064619a9b71a2e9cb72b38f8702c2e0a02ef496b7a` * **ICP Mainnet Registry**: Verification is anchored to the Internet Computer protocol ledger (when `AEGIS_NETWORK=ic`). For the FFI Rust headers and Python bindings, visit the [Zero Trust Physics GitHub repository](https://github.com/ZeroTrustPhysics/Spectrum/tree/main/ztp-dexterous-hand).