fuxi-robot/excavator-video

EXC is a comprehensive, large-scale dataset and benchmark suite curated for advancing autonomous excavator truck loading in intelligent heavy machinery and robotic earthmoving systems. It is accompanied by pre-trained perception models, standardized performance benchmarks, and a suite of simulation tools to democratize access to high-fidelity, multi-modal construction data for the academic community and the heavy equipment industry.

We have released:

• EXC Video: Our video dataset contains 10 different trucks, with each truck having 100 complete loading video trajectories, for a total of 1,000 video trajectories.
• EXC Motion: A curated subset of EXC, containing 92,214 trajectories (~xT)

Key Features 🔑

• 1000+ complete loading cycles from 10+ unique excavator.
• 200+ distinct operational scenarios across 5 key industrial domains (e.g., mining, urban construction, quarrying).
• Advanced multi-modal sensing: High-density LiDAR / 360° RGB-D camera coverage / Boom, arm & bucket IMUs / Hydraulic pressure & engine telemetry.
• Diverse and challenging loading conditions: Varying material types (soil, gravel, rock), dynamic truck positioning, confined workspaces, and adverse weather & lighting.

Day-A

Day-B

Night-A

Data Collection and Processing Workflow

The entire workflow can be divided into the following five main steps:

1. Raw Data Decryption and Conversion

   • Source Data Format: Encrypted video files in fiv format, each approximately 20 minutes long.
   • Processing Method: Use the specified decryption tool (MDVRPlayerSetupR002V0102.zip) to convert the encrypted fiv files into a common video format (e.g., MP4) for further processing.

2. Raw Video Screening

   • Objective: To screen the converted long videos and identify segments containing the "excavator loading a truck" scene.
   • Screening Criteria:
     • Scene Requirement: Must include both an excavator and a truck to be loaded.
     • Viewpoint Requirement: The excavator's hydraulic arm and bucket must be clearly visible during operation.
     • Clarity Requirement:
       • Daytime: Lens stains/obstructions should not cover more than 15% of the screen; the ore and the digging/unloading actions must be clearly visible.
       • Nighttime: Stricter standards apply. The hydraulic arm and bucket must be clearly visible. Minor frame skips are acceptable, but there should be no sudden jumps (teleportation).
     • Rejection Criteria: Videos that are obscured by dust, heavy stains, or poor angles, making it impossible to see key operational actions, must be discarded.

3. Video Clipping

   • Definition: One clip must contain one complete loading process.
   • Start Point:
     1. The excavator is ready (e.g., bucket is above the truck or empty), and the truck is either in position or entering the loading area.
   • End Point:
     1. The truck bed is filled (typically 15-20 bucket loads).
     2. The truck starts and leaves the loading area.
   • Process Requirements:
     • The bucket must be visible throughout the digging and moving process.
     • Obstruction during unloading due to the viewing angle is permissible, provided that more than 80% of the bucket is visible or the material dumping action is clearly seen.
   • Unqualified Examples: The bucket is obscured during scooping; the bucket moves out of the frame during transit.

4. Post-processing

   • Watermark Removal: Submitted video clips must have watermarks removed. Minor blurring or masking resulting from watermark removal is acceptable.
   • Format Unification: All submitted video clips must be in MP4 format.

5. Data Submission

   • Schedule: Tentatively, all processed video clips from the previous week are to be submitted every Friday afternoon.

---

Data Grading and Classification

All qualified video clips must be graded into two levels based on their quality: Class A and Class B.

• Class A Video (High Quality)

  • Image Quality: The video is clear, with no issues like dust on the lens, blurring, or poor focus.
  • Completeness: The digging and unloading points are clearly visible, and the bucket remains within the video frame throughout its movement.
  • Fluency: The total duration of non-operational actions (e.g., pauses, repositioning) during the excavator's work does not exceed 3 seconds.
  • Bucket Fill Rate: Instances where the bucket is less than 50% full do not exceed 3 times in the entire clip.
  • Stability: The camera view is stable with no significant shaking.

• Class B Video (Standard Quality)

  • Definition: Meets the basic requirements for a clip but has some defects that prevent it from meeting Class A standards.
  • Specific Issues:
    1. The video is slightly blurry, but the operational actions are still discernible.
    2. The bucket may briefly move out of the video frame.
    3. The total duration of non-operational actions (pauses, etc.) exceeds 3 seconds.
    4. The number of times the bucket is less than 50% full is 3 or more.

---

Naming Convention

All final video clip files must adhere to a strict naming convention.

• Naming Format: ExcavatorID-RecordingDate-ClipStartTime-ClipEndTime-day/night-A/B

• Field Descriptions:

  • ExcavatorID: The ID of the operating excavator.
  • RecordingDate: The date the video was recorded, in YYYYMMDD format (e.g., 20240101).
  • ClipStartTime: The start time of the clip within the original video, in HHMMSS format (e.g., 133015).
  • ClipEndTime: The end time of the clip within the original video, in HHMMSS format.
  • day/night: Indicates whether the clip was recorded during the day (day) or at night (night).
  • A/B: The quality grade of the clip (A or B).