for academic experimentation. 2. OBJECTIVES AND SCOPE The primary objective is to develop and benchmark an autonomous excavator sys- tem in a controlled dump-pocket testbed. This research addresses the “Technology” and “Transformation” themes of the conference by: (i) establishing a reproducible benchmark to quantify removal rate and autonomy reliability, and (ii) supporting progress toward reduced human exposure in hazardous confined spaces. Target KPIs are: removal rate (material removed in g/min; maximize autonomous throughput), success rate (autonomous session completion without intervention; reliability for deployment), and safety proxy (human exposure time in hazard zone; target zero). 3. METHODOLOGY OR APPROACH The system (Figure 1) is built around two SO-ARM100 robot arms [10] in a leader- follower teleoperation configuration. The SO-ARM100 is a low-cost (sub-$250), open- source, 6-DoF robot arm with 3D-printed PLA components and off-the-shelf STS3215 servo motors (7.4V), developed as an accessible platform for robotics research. The leader arm enables intuitive human demonstration via backdriving: the operator physically moves the leader, and joint angles map to the follower in real-time at 30 Hz. The follower arm executes the task equipped with a custom excavator bucket gripper. The teleoperation interface displays dual-camera streams (wrist and overhead) alongside real-time motor data to enable precise demonstration collection. This setup produced the 162-episode visuomotor dataset used for all experiments. Policy training leverages an open-source imitation learning library [9] with unified infrastructure across all evaluated architectures. 3.1 Task definition Workspace: Dump pocket testbed with internal box dimensions L × W × H = 0.454 × 0.310 × 0.112 m. Walls are vertical with inclination θ = 90◦ and maximum wall height Hmax = 0.112 m. Target region: The target region is not pre-defined as a fixed ROI; instead, the learned policy learns the cleaning objective implicitly from teleoperation demonstrations. Material: Natural bentonite clay (commercial cat litter, 0.5–2.5 mm particle size). The material exhibits granular behavior with slight interparticle cohesion, providing stable scooping dynamics. Bulk density ρbulk ≈ 1500–1700 kg/m3; testing conducted under dry indoor conditions. Initial conditions: Each run begins with 3000 g of material in the dump pocket (±20 g variation, <1%). The excavator bucket has 18 g max capacity per scoop. The evaluation measures continuous removal rate (g/min); complete pocket evacuation within the session is not required. Task cycle: The policy detects regions requiring cleaning, approaches with the bucket, scoops material (18 g max capacity), transports to the primary crusher at the center
RkJQdWJsaXNoZXIy MTM0Mzk2