IL-ACT: Imitation Learning with Adaptive Cartesian Tracking Control for a 30-ton Excavator

Mehdi Heydari Shahna; Seihun Kim; Soyi Jung; Soohyun Park; Jouni Mattila; Joongheon Kim

Source-linked AI summary

IL-ACT: Imitation Learning with Adaptive Cartesian Tracking Control for a 30-ton Excavator

Mehdi Heydari Shahna, Seihun Kim, Soyi Jung, Soohyun Park, Jouni Mattila, Joongheon Kim

arXiv:2609.16696v1Sep 15, 2026cs.RO

TL;DR

Autonomous excavator control must handle coupled kinematics, hydraulic lag, and uncertainty during goal reaching and trajectory tracking. IL-ACT combines an anchored imitation policy with adaptive Cartesian feedback, gated gain/bias estimation, and a stopping-distance governor. Across simulated goals and trajectories, it outperforms Teacher+ACT on duration, terminal error, and tracking RMSE in the reported conditions, while estimator benefits are measurable and pretrained-weight effects remain mixed.

Problem
Autonomous excavator control is challenged by coupled kinematics, actuation lag, and uncertainty during goal reaching and trajectory tracking.
Method
IL-ACT combines an anchored imitation policy pretrained from operator demonstrations with adaptive Cartesian feedback, gated gain/bias estimation, and a stopping-distance governor.
Results
IL-ACT completes all 100 goals with shorter duration and lower terminal errors than Teacher+ACT under nominal and disturbed response, while telemetry-initialized IL-ACT lowers RMSE across the expanded tracking comparisons.
Takeaways & Limitations
The evaluated simulations support combining learned nominal commands with adaptive Cartesian correction for robust excavator goal and trajectory control.
Takeaways & Limitations
Pretrained-weight effects remain mixed, estimator benefits depend on conditions, and reference admissibility is conditional on governor feasibility.

Abstract

from arXiv · show

Autonomous excavator control is challenged by coupled kinematics, actuation lag, and uncertainty. We propose imitation learning and adaptive Cartesian tracking (IL-ACT), a novel motion control framework for a 30-ton-class excavator. An anchored, 14-input imitation policy pretrained on operator demonstrations generates nominal joint rates; adaptive Cartesian feedback and gated gain/bias estimation correct these commands before a stopping-distance governor constrains joint-reference generation. Simscape evaluation covers 100 sequential goals and spiral, figure-eight, and rounded-raster tracking, including 88 additional runs across three training seeds, two initializations, and speeds, under hydraulic response and sensing conditions. Compared with Teacher+ACT, IL-ACT completes all goals with shorter duration and lower terminal errors under both response conditions. Telemetry-initialized IL-ACT lowers RMSE in all 24 figure-eight and rounded-raster seed comparisons and lowers additional-load spiral mean RMSE by approximately 29%. Original spiral RMSE also improves over IL-only and PID. Under a shared sensor-noise realization, telemetry-initialized IL-ACT achieves 27.67% lower mean RMSE than Teacher+ACT; enabling estimation reduces mean RMSE by $22.44\%$ relative to the frozen estimator. Pretrained-weight effects remain mixed, and the original teacher comparison exhibits a spiral RMSE--maximum-error tradeoff. Analysis establishes bounded adaptive states and Cartesian feedback, with reference admissibility conditional on governor feasibility.

I. INTRODUCTION

The paper targets goal-reaching and trajectory tracking for a 30-ton-class excavator by combining imitation learning with adaptive Cartesian control and constrained reference generation.

I. INTRODUCTION: Autonomous excavation spans localization, machine positioning, goal reaching, and trajectory tracking; this work focuses on the latter two control stages.The controlled machine has four joints: swing, boom, arm, and bucket.
I. INTRODUCTION: IL-ACT combines an anchored joint-rate policy with Cartesian feedback, gated gain/bias estimation, and a shared stopping-distance governor.The framework addresses coupled kinematics and hydraulic actuation while requiring reference admissibility conditional on governor feasibility.
I. INTRODUCTION: The anchored policy uses kinematic teacher supervision, dataset aggregation, causal observations, and an exact zero-action anchor for coordinated four-joint control.Telemetry-initialized and random-weight variants are evaluated.
I. INTRODUCTION: The evaluation compares IL-ACT with IL-only, tuned PID, and Teacher+ACT across goal regulation, spiral, figure-eight, rounded-raster, and additional-load tracking.The expanded study includes three training seeds, two initializations, multiple speeds, and response settings.
I. INTRODUCTION: The controller operates at 0.1 s intervals using measured joints, measured bucket-tip position, desired position and velocity, and a policy-conditioning point.Goal regulation uses the current fixed goal as the conditioning point with zero desired velocity, whereas timed tracking supplies a preview point to the learned policy.

III. IL POLICY DESIGN AND TRAINING

The IL policy is trained from operator telemetry with teacher-generated Cartesian-velocity labels and checkpoint selection based on validation success and terminal error.

III. IL POLICY DESIGN AND TRAINING: The original telemetry corpus contains demonstrations from 15 recording dates partitioned into training, validation, test, and stress-test dates.The original policy is a goal-conditioned four-joint behavior-cloning system initialized from a telemetry-trained policy.
III. IL POLICY DESIGN AND TRAINING: A kinematic teacher reconstructs configuration and Cartesian goal, selects a bounded elbow-down posture, and blends posture guidance with box-constrained Cartesian-velocity control.Teacher labels are joint-rate commands before the stateful governor.
III. IL POLICY DESIGN AND TRAINING: The initial supervised dataset contains 120000 accepted examples, and each stage uses AdamW with checkpoint selection driven first by validation success and then mean terminal error.The selected fourth-stage checkpoint is evaluated on 64 separate test cases.
III. IL POLICY DESIGN AND TRAINING: The training loss compares predicted and teacher Cartesian velocities using joint-rate scaling, Cartesian normalization, and state-dependent example weights.Its definition remains fixed while dataset aggregation changes the empirical training distribution.

IV. CONTROL DESIGN

IL-only and IL-ACT use the same learned policy, observations, tasks, and command governor, while IL-ACT additionally applies Cartesian correction and gain/bias adaptation.

IV. CONTROL DESIGN: IL-ACT enables Cartesian correction and gain/bias adaptation upstream of the governor, whereas IL-only disables both components.This isolates the contribution of adaptive tracking while holding the learned policy and evaluation conditions fixed.
IV. CONTROL DESIGN: The nominal plant response uses two cascaded 0.1 s position-input filters, while the disturbed response adds valve, friction, and supply/load effects.The comparison therefore tests controller behavior under nominal and perturbed hydraulic response conditions.

A. Design of the Goal-Conditioned Imitation Policy

The goal-conditioned policy uses causally filtered observations and a 14-input anchored network to generate four-joint nominal rate commands with an exact zero-action condition.

A. Design of the Goal-Conditioned Imitation Policy: The observer unwraps swing and causally filters measured joints and bucket-tip position before policy conditioning.A large joint or tip jump triggers an observer reset; resets after controller initialization are treated as faults.
A. Design of the Goal-Conditioned Imitation Policy: The ordered observation encodes pose and policy-conditioning error without a history stack or binary mask; desired velocity enters Cartesian feedback separately.The conditioning horizon is a time parameter, not a completion deadline.
A. Design of the Goal-Conditioned Imitation Policy: The 14-input network has 256–256–128 hidden units, SiLU activations, 103044 trainable parameters, normalized inputs, and an affine rate output.It produces swing, boom, arm, and bucket rates in deg s−1.
A. Design of the Goal-Conditioned Imitation Policy: The anchored policy shares weights and normalization across evaluations, and its nominal action is exactly zero when conditioning error and measured rate are zero.Weights are trained offline and frozen online; command limiting follows feedback and adaptation.

B. Adaptive Cartesian Tracking Correction

IL-ACT corrects learned joint-rate commands using measured Cartesian error, filtered nominal-response residuals, and gated adaptive feedback. Updates are disabled in IL-only mode and otherwise constrained by eligibility, scheduling, projection, and anti-windup rules.

Feedback correction: Measured Cartesian error, rather than policy preview error, drives the feedback correction.The controller forms e_k from desired and measured tip positions and uses the current Jacobian and its minimum singular value for conditioning.
Adaptive estimation: The estimator compares measured motion with nominal position-filter predictions to compensate gain and bias deviations in joint response.Nominal filter states are driven by the previously emitted demand, and the resulting residual supports effective gain and bias estimates.
Adaptive estimation: Gain and bias updates occur only after startup and under combined-control, excitation, residual, and non-intervention conditions; otherwise estimates remain fixed.The intervention flag is delayed to match emitted-demand timing, becoming the limiting flag from two controller samples earlier.
Gain scheduling and integration: Scheduled gains update only when the preceding sample had neither task-velocity nor governor limiting, while initial values are k_p,0 = 0.2 and ρ_0 = 0.5.Projection and scheduling constrain the adaptive quantities, and anti-windup integration uses a validity indicator that is zero under task or governor limiting.
Operating modes: IL-only mode uses the imitation command directly and disables feedback, estimator, scheduled-gain, and integral updates.This isolates the learned policy from the adaptive Cartesian correction pipeline.

C. Shared Command Governor and Execution

A shared governor converts controller commands into admissible joint references using rate, acceleration, position, and stopping-distance constraints. Execution includes fault handling, interval-aligned demand histories, and feasibility-dependent fallback behavior.

Shared limits: The governor constrains the position-demand register with shared position, speed, and acceleration limits for swing, boom, arm, and bucket.Swing has no finite position bound in this model, but its rate and acceleration remain limited.
Stopping-distance governor: Stopping-distance constraints combine the register position with the previous command and acceleration limit to bound each joint’s admissible speed.The governor uses the position-demand register rather than measured plant position, and stopping requests zero rate before acceleration and stopping-distance constraints.
Feasibility: A governor sample is feasible only when the register is finite, inside the position envelope, and each lower admissible bound does not exceed its upper bound.Reference admissibility therefore depends on feasibility of the governor inequalities at each sample.
Fault handling: Invalid inputs, observer faults, inconsistent startup, or infeasible admissible intervals trigger latched faults that reduce commands toward zero through limited deceleration.The initialization sample holds the initial demand and emits zero rate; infeasibility uses a fallback reduction toward zero.
Execution: The controller emits the current demand register while storing the next register with interval-aligned histories, and IL-only omits feedback and adaptation.Stopping and governor processing precede anti-windup integration in the valid-sample execution sequence.

D. Controller Boundedness and Reference Admissibility

The analysis bounds the adaptive states and Cartesian feedback, then proves that initially admissible generated references remain admissible while governor feasibility is maintained. The combined command is bounded when the learned nominal output is bounded on the operating domain.

Analytical guarantees: The projected parameters satisfy 0.35 ≤ b_gj,k ≤ 1.8, |b_bj,k| ≤ 0.12, 0.2 ≤ k_p,k ≤ 0.6, 0.5 ≤ ρ_k ≤ 10, and ∥I_k∥_2 ≤ 25.The analysis also bounds the pseudoinverse-related term and Cartesian feedback command.
Command boundedness: If the learned nominal output is bounded on the operating domain, the combined command is also bounded because estimated gains remain strictly positive.This conclusion relies on the parameter bounds and the bounded feedback command.
Reference admissibility: An initially admissible position-demand register remains admissible while the governor remains feasible because discrete stopping distance includes the next commanded displacement.The stopping-distance construction supplies the key step in the admissibility argument.
Analytical guarantees: Bounded adaptive states, Cartesian feedback, and generated references are established under the stated governor-feasibility conditions.The result is conditional: admissibility is not asserted when the governor’s feasibility requirements fail.

A. Original Training and Analytical Rollouts

The original evaluation tests IL-ACT on sequential Cartesian goals and spiral tracking under nominal and disturbed hydraulic response. IL-ACT outperforms Teacher+ACT on completion duration and terminal error while maintaining bounded, governor-constrained operation.

Training and validation: The original single-seed policy passes all 32 validation cases at each of four stages, reducing mean terminal error from 21.431 to 20.367 mm.Success requires 25 mm position and 0.05 deg s−1 speed tolerances, with qualification, hold, and timeout rules.
Validation checks: Across original IL-only and IL-ACT timed runs, all eight runs avoid controller faults and joint-limit violations, while disturbed IL-ACT alone performs gain/bias updates.The runs include four goal-regulation and four spiral cases; the largest reported update counts are 11192 for goal regulation and 10014 for spiral tracking.
Control design: IL-ACT combines Cartesian feedback and gated gain/bias estimation with a shared stopping-distance governor that constrains generated joint references.The evaluation compares against a Cartesian-space PID baseline whose gains are tuned separately and then frozen.
Goal regulation: Both controllers complete all 100 sequential goals, while IL-ACT reduces duration by 7.22% nominally and 12.97% under disturbance, with terminal-error reductions of approximately 16.16% and 28.39%.The comparison uses shared acceptance and governor rules; enabling estimation additionally lowers disturbed-goal mean terminal error by 12.68%.
Spiral tracking: IL-ACT’s original spiral comparison averages 27.88% lower RMSE than Teacher+ACT but incurs 33.18% higher maximum error.This establishes a spiral RMSE–maximum-error tradeoff rather than uniform improvement across metrics.

C. Expanded Tracking and Initialization Study

The expanded study evaluates learned policies across held-out paths, temporal rates, hydraulic-response conditions, seeds, initializations, and estimator settings. Telemetry-initialized IL-ACT consistently improves RMSE over Teacher+ACT in the reported comparisons, while estimator and pretraining effects vary by condition.

Evaluation design: The study averages learned entries across seeds 11, 22, and 33 under nominal and hydraulic-response conditions, with paths evaluated at 1× and 2× temporal rates.The figure-eight and rounded-raster paths are held out from training, tuning, and checkpoint selection.
Estimator and initialization effects: 27.67% lower mean RMSE than Teacher+ACT is achieved by telemetry-initialized IL-ACT under the shared sensor-noise realization.The same study reports 22.44% lower mean noise RMSE and 17.5% lower mean per-run tracking maximum when estimation is enabled.
Evaluation validity: All 88 runs completed without recorded controller faults or operating-limit violations, and the largest tracking-window error was 12.515 mm against a 25 mm criterion.The evaluation used a 600 s approach and a 30 s terminal observation.

VI. DISCUSSION AND CONCLUSIONS

Across the evaluated simulation conditions, IL-ACT improves selected goal-regulation and tracking outcomes relative to IL-only and Teacher+ACT. Its adaptive analysis establishes bounded states and feedback properties, while several effects and tradeoffs remain condition-dependent.

Overall findings: IL-ACT improves goal-regulation duration and spiral RMSE over IL-only, and goal duration and terminal accuracy over Teacher+ACT.Lower RMSE versus Teacher+ACT also extends to figure-eight and rounded-raster paths at both speeds and response settings, plus additional-load spiral tracking.
Caveats and tradeoffs: Pretrained-weight effects remain mixed, and estimator benefits depend on conditions.The original comparison also includes higher joint command-tracking RMSE and disturbed-goal terminal error than IL-only.
Analysis and scope: The analysis establishes bounded adaptive states and Cartesian feedback, with reference admissibility conditional on governor feasibility.These conclusions concern the evaluated simulation conditions.