_{MASTER'S RESEARCH · ROBOTICS · 2026}

Can a hydraulic arm
learn to adapt while it works?

Learning-Based Online Adaptation for Control of a Hydraulic Manipulator
Kevin Kudakwashe Murera · Robotics Research Lab · RPTU Kaiserslautern-Landau

A simulated MULAG verge mower must keep its cutter on a changing path while terrain and payload disturb the arm. This project asks what learned actuator models and an online payload estimate actually contribute to closed-loop control.

Read the thesis  ·  Explore the experiments  ·  Reproduce a run

_{Simulation evidence throughout. This repository makes no claim about deployment on hardware.}

Front-view rollout · Back-view rollout · Video provenance

_{The front and back videos are separate runs of the same adaptive arm, not two synchronized cameras.}

---

The verge mowing challenge

Tracking is the objective. Terrain is the disturbance. The truck moves along the road while a hydraulic arm reaches over the embankment. Its cutter must follow a target line and stay near the ground as the terrain changes. The R and IF policies are evaluated over a 26 m path, 11 trajectory families, 2–7 km/h, and three terrain meshes. Task and pipeline overview.

The research follows two questions:

1. How does the actuator representation affect closed-loop control when the task and scoring conditions are held fixed?
2. When does estimating payload online help a frozen controller, and when does it cost more than it gains?

Research principles

Principle	What it means in this project
Pair the conditions	Compare the same mesh and trajectory family. Report paired differences, sign counts, and paired bootstrap intervals.
Verify the plant	Confirm that a requested payload or other perturbation reached the simulated plant before interpreting a controller response.
Validate the instrument	Name the checkpoint and scoring horizon, and measure whether an apparent effect exceeds the instrument's variation.
Keep claims bounded	Do not compare metrics across collections with different state distributions or normalisers. Separate model prediction, closed-loop performance, and hardware claims.

These rules, including corrections to earlier interpretations, are recorded in the experiment index.

Three routes from policy to plant

Stack	Command path	Role in the study
R · PPO-ROM	Policy → command ROM → PhysX	Reduced-order actuator path.
IF · PPO-ID-FD	Policy → inverse model → forward model → PhysX	Matched task with a different actuator path.
AF · PPO-AAM-FD	Policy → valve command → frozen forward model → PhysX	Within-stack online adaptation against its own frozen base.

R and IF choose a cylinder increment; AF chooses a valve command. AF is therefore not ranked against R and IF on task error. The individual stack READMEs give the runnable details.

What the experiments show

01 · Adaptation helps when the frozen arm has a heavy-payload deficit

On held-out mesh 3141, a three-seed context-training experiment compares the adaptive AF policy with the same frozen base. The base weights remain fixed; only the context input weights are trained. At 4× and 5× payload, pooled paired medians improve cut-height error, lateral error, and ground contact.

Payload	Cut-height error improvement	Lateral error improvement	Ground-contact gain
4×	71.52 mm	71.58 mm	+0.12
5×	82.18 mm	107.05 mm	+0.09

Positive values mean adaptive is better. These are paired median differences across environments pooled over three context-training seeds on one shared frozen base, not three independent base policies. The pooled 95% intervals exclude zero, but two of three seeds carry the heavy-load benefit; seed 23 significantly worsens cut height at 5×. Full result and intervals.

02 · Adaptation also has a boundary

At 3× payload, the frozen policy already maintains ground contact. Adaptation makes pooled lateral error 22.75 mm worse (95% paired interval: 17.27–30.04 mm worse), without a supported pooled cut-height gain. The result suggests gating adaptation to conditions where the frozen policy has a measurable deficit; that gate is a next step, not a demonstrated controller. Three-seed report.

03 · Both published actuator routes clear the gate suite

The promoted R and IF policies each pass all eight task gates at 2–7 km/h on meshes 1337, 17, and 909. This is a pass result for each route, not a claim that one route outperforms the other. R protocol · IF protocol.

04 · The hydraulic plant has measurable latency

Command-to-differential-pressure onset was measured at 117 ms, or seven 60 Hz control steps, across three of four amplitude tiers. That transport delay is identifiable even though the valve's full dynamic model is not identifiable from the same data. Plant characterisation and method notes.

Follow the evidence

Start here	What you will find
Full thesis	The research question, model and controller design, experiments, and limits.
Experiment index	Methods, paired-comparison rules, corrections, and Chapter 7 evidence map.
Final adaptation report	Three seeds, paired intervals, failure cases, and explicit limitations.
Pipeline READMEs	Commands and checkpoints for each stack; follow the R, IF, and AF links above.
Reproduction guide	Setup, a shipped-policy viewer, preflight, and scoring entry points.

Citation

@mastersthesis{murera2026online,
  author = {Kevin Kudakwashe Murera},
  title  = {Learning-Based Online Adaptation for Control of a Hydraulic Manipulator},
  school = {Robotics Research Lab, Department of Computer Science, University of Kaiserslautern-Landau},
  year   = {2026},
  type   = {Master Thesis}
}