智源 FlagOpen 开源 RoboCOIN-RTML：机器人轨迹质量评估 DSL 与 Robot Data Clinic

FlagOpen; RoboCOIN-RTML

Repository navigation

NotificationsYou must be signed in to change notification settings; Fork3 (3); Star1 (1)You must be signed in to star a repository

Repository files navigation

README

🤖 Robot Data Clinic (RDC)

To address the lack of a unified quality evaluation standard for robot manipulation data, we propose Robot Trajectory Markup Language (RTML) — a domain-specific language (DSL) designed for trajectory quality assessment.

RTML translates empirical quality dimensions (e.g., smoothness, stability, coordination) into computable, configurable, and machine-readable rules, supporting both global constraints and subtask-level customization.

Built upon RTML, we introduce Robot Data Clinic (RDC) — an end-to-end system for trajectory data evaluation and enhancement, forming a closed-loop pipeline: Diagnose → Filter → Enhance.

🔍 1. Trajectory Analyser

Takes raw trajectory data as input and outputs multidimensional statistical features (pose, velocity, acceleration, time, etc.); Generates a "Robot Data X-Ray", visualizing spatial distributions, orientation anomalies, and other patterns of the end-effector; Establishes an objective quality baseline for subsequent filtering

✅ 2. Trajectory Selector

Automatically filters low-quality trajectories using an RTML rule library; Supports two levels of constraints:

Global constraints: Basic quality requirements applicable to all tasks (e.g., max velocity, acceleration); Subtask constraints: Fine-grained rules tailored to specific operation phases (e.g., workspace boundaries, idle-arm motion limits); Filters rigorously across five quality dimensions:

Motion smoothness; Pose stability; Bimanual coordination; Task focus; Execution efficiency

🚀 Quick Start

Environment Setup

📊 robot_trajectory_analyser

Generate X-Ray Statistics

Compute multidimensional statistics based on CoRobot-annotated subtasks and generate per-episode X-Ray files:

💡 You can also interactively debug components and visualize end-effector pose changes and spatial trajectories in the notebook: compute_subtask_stats.ipynb.

Multidimensional Visualization

Visualize a specific statistic for a given subtask:

Batch-generate visualizations for all subtasks and statistics:

Example outputs:

Workspace Dimension Velocity Dimension

🧪 robot_trajectory_selector

Write an RTML File

Craft RTML rules based on statistical insights and subtask characteristics.

Define global constraints (task-agnostic quality baselines):

Define stage-specific constraints per subtask (e.g., workspace bounds, idle-arm limits):

📄 Full example: RTML_release/src/robot_trajectory_selector/RTML_demo.yaml

Filter High-Quality Episodes

Apply RTML rules to filter high-quality trajectories: