Technical Summary: IL-ACT for a 30-ton Excavator

Problem Statement

Autonomous control of heavy hydraulic machinery, specifically 30-ton-class excavators, faces significant challenges due to coupled kinematics, actuation lags, and system uncertainties. While previous systems have demonstrated trajectory optimization and learned inverse control, achieving robust, high-precision tracking under uncertain dynamics and changing operating conditions remains difficult. The authors address the need for a motion control framework that can handle goal-reaching and trajectory tracking for such complex machines, balancing the benefits of imitation learning (IL) with the robustness of adaptive control.

Methodology: The IL-ACT Framework

The paper proposes IL-ACT (Imitation Learning with Adaptive Cartesian Tracking), a novel motion control framework designed for a 30-ton hydraulic excavator with four controlled joints (swing, boom, arm, and bucket). The system operates in a closed loop with a 0.1s control period and consists of three primary components:

1. Anchored Imitation Policy

• Architecture: A fully connected neural network (14 inputs, 4 outputs) trained via behavior cloning on operator demonstrations.
• Training Strategy: The policy is pretrained on a telemetry corpus (15 recording dates) and refined using offline kinematic supervision. A "kinematic teacher" reconstructs configurations and generates bounded inverse-kinematic (IK) posture guides and Cartesian velocity commands.
• Anchoring Mechanism: To ensure stability and zero-action consistency, the policy output is defined as the difference between the network's prediction for the current state and its prediction for a "zero-action" anchor state (where the goal matches the current pose and velocity is zero). This ensures that when the system is at the goal with no motion, the nominal command is exactly zero.
• Observations: The input vector includes causal, filtered joint angles, tip position, and the error vector to a conditioning point (goal or preview point).

2. Adaptive Cartesian Tracking (ACT)

• Feedback Correction: The system employs a Cartesian feedback loop that calculates the error between the desired and measured tip positions.
• Gain/Bias Estimation: A gated estimator compares measured joint rates against a nominal plant response prediction. It estimates gain and bias corrections to compensate for deviations caused by hydraulic uncertainties (valve lag, friction, load modulation).
• Adaptation Logic: Updates to the gain and bias estimates are "gated" based on excitation levels (to avoid adaptation during low-signal or saturated conditions) and intervention flags.
• Integration: The final command is a combination of the nominal IL output, the Cartesian feedback velocity (mapped to joint space via a damped inverse Jacobian), and the estimated bias/gain corrections.

3. Shared Command Governor

• Safety and Feasibility: A stopping-distance governor constrains the generated joint references to ensure the machine can stop within its physical limits (position, velocity, and acceleration bounds) before reaching a target.
• Reference Admissibility: The governor ensures that the position-demand register remains within the declared envelope. If the governor detects an infeasible state, it triggers a fallback deceleration rather than an instantaneous stop.
• Anti-windup: The system utilizes anti-windup integration for the integral term in the feedback loop to prevent performance degradation during saturation.

Key Contributions

The authors identify three main contributions:

1. Anchored Imitation-Policy Design: A coordinated four-joint policy refined via dataset aggregation and kinematic supervision, featuring an exact zero-action anchor to ensure stability.
2. Adaptive Tracking with Constrained Reference Generation: A framework combining Cartesian feedback and gated gain/bias estimation with a shared stopping-distance governor. The paper provides a theoretical analysis establishing the boundedness of adaptive states and the admissibility of generated references, conditional on governor feasibility.
3. Comprehensive Comparative Simulation Evidence: Extensive evaluation across multiple tasks (goal regulation, spiral, figure-eight, and rounded-raster tracking) and conditions (nominal vs. disturbed hydraulic response, sensor noise, additional load). The study includes comparisons against tuned PID, IL-only, and Teacher+ACT baselines, utilizing multiple training seeds and initialization strategies.

Experimental Results

The framework was evaluated in a high-fidelity Simscape environment simulating a 30-ton excavator with hydraulic disturbances (valve lag, friction, hysteresis, load modulation).

Goal Regulation (100 Sequential Goals)

• Success Rate: Both IL-only and IL-ACT achieved 100/100 success in both nominal and disturbed conditions, whereas PID achieved 95/100 (nominal) and 86/100 (disturbed).
• Efficiency: Compared with Teacher+ACT, IL-ACT completes all goals with shorter duration and lower terminal errors. Specifically, IL-ACT reduces duration by 7.22% (nominal) and 12.97% (disturbed) compared to Teacher+ACT.
• Accuracy: Under nominal and disturbed response, IL-ACT reduces mean terminal error by approximately 16.16% and 28.39%, respectively, compared to Teacher+ACT.

Trajectory Tracking (Spiral, Figure-Eight, Raster)

• Spiral Tracking: IL-ACT significantly outperformed both PID and IL-only. In the presence of hydraulic disturbances, IL-ACT achieved a Cartesian tracking RMSE of 0.05864 mm, compared to 12.48 mm for PID and 2.49 mm for IL-only.
• Generalization: In an expanded study involving 88 runs across three training seeds and two initializations (telemetry vs. random), telemetry-initialized IL-ACT lowered RMSE in all 24 figure-eight and rounded-raster seed comparisons against Teacher+ACT.
• Load and Noise Robustness: Under additional-load spiral conditions, telemetry-initialized IL-ACT lowered the mean RMSE by approximately 29% compared to Teacher+ACT. Under shared sensor-noise realizations, it achieved a 27.67% lower mean RMSE than Teacher+ACT.
• Estimator Impact: Enabling the gain/bias estimator reduced the mean RMSE by 22.44% relative to a frozen estimator in sensor-noise conditions.

Significance and Claims

The paper claims that IL-ACT successfully integrates the data-driven efficiency of imitation learning with the robustness of adaptive control theory.

• Robustness: The framework demonstrates superior performance over pure imitation learning (IL-only) and model-based baselines (PID) in the presence of significant hydraulic uncertainties and external disturbances.
• Theoretical Guarantees: Unlike many black-box learning approaches, the authors provide an analysis establishing that the adaptive states and Cartesian feedback commands remain bounded, and that the generated references are admissible provided the governor remains feasible.
• Practical Applicability: The results suggest that combining a pretrained policy with online adaptation and a safety governor is a viable path toward autonomous control for heavy machinery, offering a balance between the sample efficiency of learning and the safety requirements of industrial operation.

The authors remain modest regarding the scope, noting that these findings are specific to the evaluated simulation conditions and that the effects of pretrained weights can be mixed depending on the specific task and initialization. The work does not claim to have solved all aspects of autonomous excavation (e.g., target localization is noted as a separate stage), but rather focuses on the motion control and trajectory tracking stages.