Robots built from soft or "smart" materials, such as pneumatic muscles, piezoelectric stages and shape-memory wires, have a property called hysteresis. The same command does not always produce the same motion, because the output depends on the path of earlier commands, much as a bent paperclip remembers how it was bent. Meanwhile, robot manipulation is increasingly run by teams of language-model agents. A planner breaks a task into steps, a coder turns the steps into motion commands, and a supervisor watches the camera and triggers re-planning when something goes wrong. These agent teams assume that a command to move to a point takes the arm to that point. On a hysteretic actuator that assumption fails quietly. The arm lands a few millimetres off in a direction set by where it came from. The supervisor sees a miss and calls it a task failure, so the planner re-plans a task that never needed re-planning.
This note proposes a narrow fix that can be tested and could turn out to be unnecessary. The low-level hysteresis model already runs on the controller and already tracks the actuator's internal memory. We expose it to the agents as a cheap query tool and use it in two places. The first is the planner, which can choose the order of sub-goals and the direction of approach so that leftover hysteresis error is small. The second is the supervisor, which can check whether an observed deviation is what the actuator's history predicts or a real failure such as a dropped object or a collision. The claim is that this coupling reaches higher task success with fewer wasted re-planning cycles than two alternatives. One is a better compensator hidden entirely beneath the planner. The other is the obvious heuristic of resetting the actuator before every move.
Core Claim
A well-identified inverse compensator removes most hysteresis error, but not all of it. In the strongest hybrid result in the file, feedforward plus PID feedback on a 2D piezo stage still leaves errors within ±0.5 µm, with an average absolute error of 0.1192 µm (Wang et al.). In the pneumatic cylinder study, force error remains below 0.5 N with an overall RMS of about 1 N, and transient spikes occur at direction reversals (Wu et al.).
This leftover error is not random. It depends on approach direction and on the recent loading history. Both of these are set by decisions the planner makes: sub-goal order, via-points, and whether to retreat before re-approaching. Our hypothesis is that in multi-step tasks with tight placement tolerances, hysteresis-aware choices at the planning level and hysteresis-aware failure diagnosis at the supervision level each add task success beyond what the best hidden compensator delivers, and that the gain grows with the hysteresis amplitude left after compensation.
The supervisor's test is simple. Let $\hat{y}(h)$ be the end-point the hysteresis model predicts given the commanded path history $h$, let $\sigma$ be the model's calibrated residual spread, and let $y_{\text{obs}}$ be the observed end-point. The supervisor treats a deviation as an actuator-memory effect, to be handled by a local re-approach, when
$$|y_{\text{obs}} - \hat{y}(h)| \le k\sigma,$$
and as a genuine task failure, to be escalated to the planner, otherwise. Here $k$ is a threshold tuned on held-out episodes.
Relation to Prior Work
The agent-team side supplies the architecture we modify. MALMM splits manipulation into Planner, Coder and Supervisor agents on RLBench. It reports an average success of 0.81 across 9 tasks against 0.70 for the best baseline, and more than 6x better success than a single agent on four-block stacking. It assumes ideal kinematics, and we keep its structure while removing that assumption. Agents4PLC shows a design point we rely on: agents generate and verify code offline, and deterministic controllers execute it in real time. That is how we keep slow language-model turns out of the 2 ms control loop.
The hysteresis side supplies the model the agents query. Chandra et al. use sparse equation discovery (SINDy, a regression that picks a few terms from a library of candidate functions) to fit a 5-term differential equation to 45,000 real piezo cantilever samples. It reaches $R^2 = 0.99$ and trains in 0.018 s, which makes it cheap enough to roll forward thousands of times per planning query. Their own reported sensitivity to outliers is a constraint: we must calibrate $\sigma$ on held-out data rather than trust the fit. Operator models such as the multi-slope Prandtl-Ishlinskii model of Wang et al. provide a second, structurally different model, so the result does not hinge on one model family.
The closest bridging paper, Lin et al., compensates Bouc-Wen hysteresis in actuators and sensors across a consensus network of dynamical agents. "Agents" there means synchronized control loops, not role-specialized planners, and the hysteresis never reaches task-level decisions. Our work differs in exactly that respect.
Two papers motivate the problem. ManiSoft shows that end-to-end policies on a soft continuum arm collapse: Diffusion Policy falls from 31.6% to 2.2% between clean and randomized scenes, and OpenVLA reaches 27.0% on randomized scenes. The musculoskeletal controller of Driess et al. works only for movements that start from a fixed resting state, because hysteresis breaks its history-free model. That restriction is in effect the "reset before every move" heuristic, and we treat it as a baseline rather than as a solution.
Research Program
Testbeds. We build two simulated testbeds and one physical check.
The first testbed adds hysteresis to RLBench (RLBench). A hysteresis operator sits between commanded and realized joint or end-effector positions on the Franka arm. Its parameters are fitted to the Chandra cantilever data and then scaled to a sweep of amplitudes. We run the same 9 tasks MALMM used, plus tighter-tolerance variants of stacking and insertion, where millimetre-level leftover error decides success.
The second testbed applies the same treatment to ManiSoft's Cosserat-rod arm. A hysteresis operator is placed on each actuator input, so the internal memory is real but invisible to the camera.
The physical check is a single-acting pneumatic cylinder bench like Wu et al.'s, driving a one-degree-of-freedom pick-and-place stage. It tests whether the supervisor's diagnosis rule holds up under real sensor noise.
System. On the low-level controller, an inverse compensator (SINDy-based or multi-slope Prandtl-Ishlinskii, plus PID) runs at control rate. This compensator is shared by all hysteresis-aware conditions, so the comparisons isolate what the agents do with the model.
The same forward model is wrapped as two tools. The first is predict_endpoint(path), which returns $\hat{y}$ and $\sigma$ for a candidate command sequence starting from the current memory state. The second is memory_state(), which returns a compact summary of recent loading history.
The planner prompt asks it to compare at most a few candidate sub-goal orderings or approach directions through the first tool. The supervisor applies the test above before escalating. No language-model call is ever placed inside the control loop.
Baselines
The naive reference point is unmodified MALMM running on the hysteresis-injected testbeds with no compensator. It shows how much hysteresis hurts an agent team that assumes ideal execution. Its nominal-condition number is the reported 0.81 average RLBench success (MALMM). We found no reported result on Lacuna for MALMM under actuator hysteresis, so we measure it ourselves.
The strongest competitor is MALMM on top of the best hidden compensator: the same SINDy or multi-slope inverse with PID that our system uses, invisible to the agents. This is the design a controls engineer would propose. The closest reported figures for this layer alone are Wang et al.'s ±0.5 µm hybrid error (average absolute 0.1192 µm) and Wu et al.'s sub-0.5 N force error. These are actuator-level numbers, and we found no reported result on Lacuna for task success with such a compensator under an agent planner.
The heuristic baselines come next. The first is "reset before every sub-goal": retract or depressurize to a rest state, as the musculoskeletal work effectively does. The second is "always approach from the same side", the classic machinist's trick for backlash.
The monolithic reference on ManiSoft is OpenVLA and Diffusion Policy at their reported 27.0% and 2.2% randomized-scene success. We also run a variant that is fed proprioceptive command history as input.
Fairness rules apply to every agent condition. All conditions share the same language model, prompt budget, cap on re-planning turns, and per-episode wall-clock limit. All compensators share identical identification data. The heuristics receive the same compensator.
Two ablations isolate the contribution: hysteresis-aware planning only, and hysteresis-aware supervision only. Beating the strongest baseline means two things. Success must be higher on the tight-tolerance tasks, at equal or lower wall-clock time. Spurious re-plans, defined as re-plans triggered when no task failure occurred and identified by simulator ground truth, must also be fewer. Both differences must hold with confidence intervals across at least three hysteresis amplitudes.
Why not just…?
Why not just use a better compensator and leave the planner alone? This is the most serious threat. If leftover error after compensation is below task tolerance, our system should add nothing, and the amplitude sweep is designed to show exactly that. The prediction is a crossover: no gain at low leftover amplitude, and growing gain once leftover error approaches the placement tolerance. Wang et al. and Wu et al. both report nonzero leftover error that depends on direction, which suggests the crossover exists. If it does not appear at realistic amplitudes, we will report that the compensator alone is sufficient.
Why not just reset the actuator before every move? Resetting erases memory, so it should match our accuracy. It costs time, though, and on soft arms a retract-and-return path can sweep through obstacles. We will plot success against episode time for the reset heuristic, same-side approach, and our planner. The heuristic wins if it reaches the same success within comparable time. We think that outcome is plausible on simple RLBench tasks and less likely in cluttered ManiSoft scenes.
Why not just give the supervisor a wider tolerance? A fixed tolerance cannot tell a 3 mm lag that the actuator's history predicts from a 3 mm slip caused by a bad grasp. Widening the tolerance hides real failures, and narrowing it triggers spurious re-plans. We tune a fixed-tolerance supervisor on the same held-out episodes and compare both detection rates. The prediction-based rule should dominate wherever the predicted deviation varies across episodes.
Why not just train an end-to-end policy with command history as input? ManiSoft's collapse of monolithic policies under randomization is evidence against this, but those policies lacked history input. We include that variant so the question is settled directly rather than by citation.
Decisive Checks
The idea fails in two cases. One is if the hidden-compensator baseline matches full success at every amplitude. The other is if the reset heuristic matches it at equal time. In either case the agents gain nothing from knowing the actuator's memory, and we will report it.
The idea holds if three things happen together. The planning-only ablation must improve tight-tolerance success. The supervision-only ablation must reduce spurious re-plans without missing more injected real failures, such as dropped objects, which we plant in a fixed fraction of episodes. Finally, the supervisor's diagnosis rule must keep its accuracy on the physical pneumatic bench, where the model's leftover errors are not the clean ones we simulated.
References
- Hysteresis Characteristics and MPI Compensation of Two-Dimensional Piezoelectric Positioning Stage
- Data-Driven Feedforward Force Control of a Single-Acting Pneumatic Cylinder with a Nonlinear Hysteresis Characteristic
- MALMM: Multi-Agent Large Language Models for Zero-Shot Robotic Manipulation
- Agents4PLC: Automating Closed-loop PLC Code Generation and Verification in Industrial Control Systems using LLM-based Agents
- Discovery of sparse hysteresis models for piezoelectric materials
- Adaptive Neural Consensus Tracking Control for Multi-agent Systems with Unknown State and Input Hysteresis
- ManiSoft: Towards Vision-Language Manipulation for Soft Continuum Robotics
- Learning to Control Redundant Musculoskeletal Systems with Neural Networks and SQP: Exploiting Muscle Properties
- RLBench: The Robot Learning Benchmark & Learning Environment.