← Back to aryangoyal7.github.io

Measuring the Stability Assumption Behind Action Chunking

Independent Researcher
Per-state measurement of open-loop and closed-loop error propagation: a perturbation at state s_t is followed under open-loop replay of the recorded actions and under closed-loop replanning by the deployed policy; the log state deviation is fitted to obtain a propagation rate that is classified as stable, deadband, or unstable
At a state \(s_t\), a small action perturbation \(\delta\) produces \(\tilde a_t = a_t + \delta\). We measure the deviation from a nominal rollout under two regimes: (a) open-loop, where the rest of the recorded chunk is replayed without replanning, and (b) closed-loop, where the deployed policy replans at each subsequent step. For each regime, a line fitted to the log deviation gives a propagation rate \(\lambda_{\mathrm{ol}}\) or \(\lambda_{\mathrm{cl}}\), classified as stable, deadband, or unstable.

Abstract

Action chunking improves the performance of policies learned by behavioural cloning, and several mechanisms have been proposed to explain why, including temporal consistency, horizon reduction, representation learning, and reduced error compounding. Every one concerns the error the policy introduces, whether by reducing how much of it enters or how often it re-enters the policy's own input. We instead study what happens to an action error once it enters the system. At each state, we inject a small action error and measure how fast it grows or shrinks under two execution regimes: open-loop, where the rest of the chunk is replayed without replanning, and closed-loop, where the policy replans after the perturbation. The fitted rate labels each state as contracting, expanding, or unresolved. Across twelve manipulation tasks from three benchmark suites, we find that confidently stable states are rare, while error amplification is common among states whose propagation rate can be resolved. We further find that the measured propagation rate depends strongly on the fitting horizon: amplification is typically front-loaded, so short windows can substantially overestimate longer-horizon propagation. Finally, we train predictors on these labels and find that a state's open-loop regime can be recovered from camera frames and proprioception alone, without a simulator, while its closed-loop propagation is only partially recoverable because it also depends on how the policy acts after the perturbation. These results suggest that error-compounding arguments alone do not provide a complete account of action chunking: neither passive open-loop dynamics nor policy replanning consistently contracts an injected error, and replanning rarely turns open-loop amplification into confident contraction. This also suggests that closed-loop reactivity should be trained explicitly, using perturbation- and tree-coverage-oriented training to expose the policy to deviations it must recover from, rather than expected to emerge reliably from standard imitation learning.

Background

Why is action chunking said to work?

Imitation-learned manipulation policies increasingly predict chunks of actions rather than single steps. A chunked executor plays \(k\) actions open-loop and then replans, and \(k\) is chosen once per task. Chunked execution is widely reported to outperform single-step execution, but the explanation is contested. The accounts on offer attribute the gain to different mechanisms:

  • Temporal consistency. Modelling actions jointly matches correlated demonstrator behaviour that independent sampling does not.
  • Horizon reduction. Suboptimality scales with the number of decisions, and a chunk makes \(k\) times fewer.
  • Representation learning. The harder training target yields a better internal representation.
  • Non-Markovian expressivity. A chunk can capture history-dependent demonstrator behaviour that a Markovian policy cannot.
  • Delayed conditioning. Predicting later actions from earlier observations can reduce the effect of the policy's own accumulated errors.
  • Implicit ensembling. A chunked policy learns multiple temporal relationships between observations and actions, producing ensemble-like robustness.
  • Limited context. Chunking substitutes for history the policy cannot condition on, and the advantage recedes once the context window is extended.

Every one of these accounts concerns the error that enters the loop: how often the policy introduces one, or how large it is when introduced. None of them addresses what happens to the error afterwards, whether the dynamics contract it or expand it, and small deviations are inevitable in learned policies.

The theory

The guarantees assume a single contraction rate

The theoretical bounds do speak to what happens after an error enters, but they assume a single global contraction rate for the dynamics rather than a rate that varies with state. Zhang et al. (2025) assume the true dynamics are open-loop exponentially incrementally input-to-state stable with a fixed decay factor \(\rho\in(0,1)\): roughly, \(\rho\) determines how quickly a deviation between trajectories decays from one step to the next. Under this assumption they show that sufficiently long action chunks restore exponential stability of the learned closed loop, with the required chunk length

\[k \;\ge\; k^{\star} \;=\; \log(1/\rho)^{-1}\,\log\!\big(\mathrm{poly}(L, C_{\mathrm{ISS}})\big),\]

so weaker contraction (larger \(\rho\)) requires longer chunks. Lazzati et al. (2026) likewise derive their bound under deterministic Lipschitz dynamics. In both analyses the relevant dynamical constants are assumed rather than measured and treated as global. A manipulator in free space and the same manipulator during contact need not propagate an error at the same rate.

Replanning is the other way an error can be dealt with, and it raises the same question about the deployed policy rather than the plant. The expert paired with the true dynamics, \((\pi^{\star}, f)\), is stable by construction, since a human demonstrator sees the error and corrects it. What matters at deployment is the learned policy paired with the same dynamics, \((\hat{\pi}, f)\), and stability of the former does not imply stability of the latter. Whether \((\hat{\pi}, f)\) is stable at a given state is an empirical property of the trained policy in its environment, to be measured rather than inherited from the demonstrator.

Our framework

Measuring open- and closed-loop stability per state

The idea is a small experiment repeated at many states. Pick a state, nudge the first action a little, and watch how far the robot drifts from where it would have gone without the nudge. If the gap shrinks over the next few steps, that state is stable; if it grows, it is unstable.

  1. Nudge. At state \(s_t\), add Gaussian noise to the translation part of the action. The noise scale \(\sigma_u\) is the trained policy's own action error on held-out demonstrations, so the nudge is the size of a mistake the policy actually makes. Do this \(N{=}8\) times to get eight perturbed branches.
  2. Roll forward, two ways. Open-loop: after the nudge, replay the recorded actions unchanged. No policy is involved, so this measures what the physics and the task geometry do with the error. Closed-loop: after the nudge, let the deployed policy replan at every step. Nominal and perturbed branches share the same random seeds, so any difference comes from the nudge, not from the policy sampling differently.
  3. Fit a rate. Over a window of \(K\) steps, fit a line to the log of the gap between each perturbed branch and the nominal one, \(\log \lVert y_j - y_j^{(n)} \rVert \approx \alpha_n + \lambda_n j\). The slope is the propagation rate: positive means the gap grows, negative means it shrinks. Average the eight slopes into \(\bar{\lambda}\) with a standard error.
  4. Label. Two checks. Is the rate big enough to matter, at least an \(m\)-fold change over the window (\(|\bar{\lambda}| \ge \ln(m)/K\))? And is its sign trustworthy (\(|\bar{\lambda}| \ge t\,\mathrm{SE}\))? Pass both and the state is stable or unstable by the sign; fail either and it goes in the deadband, which means "we can't tell", not "the rate is zero". Defaults: \(m{=}2\), \(t{=}2.365\), \(K{=}24\).

The open-loop probe is run at three kinds of states, and the closed-loop probe at the policy's own replan points, so the two regimes can be compared at identical states.

Stamp typeState fromActions used\(K\)
demonstrationdemonstration filedemonstrator actions, as recorded24
rolloutpolicy rolloutpolicy actions, as recorded24
policy chunkpolicy rollout, at a replan boundarychunk emitted by the policy8 or 24
closed looppolicy rollout, at a replan boundarypolicy replans at every step24

Demonstration versus rollout changes the state distribution; rollout versus policy chunk changes the action source. For policy-chunk stamps, \(K\) is capped by the emitted chunk length: 8 actions for the diffusion policies and up to 24 on RoboCasa.

We label twelve tasks across three benchmark families: robomimic (lift, can, square, tool_hang, per-task diffusion policies), MimicGen (coffee, nut_assembly, square_d0, stack_d0), and RoboCasa (OpenCabinet, OpenDrawer, PickPlaceCounterToCabinet, TurnOnSinkFaucet, driven by the public GR00T N1.5 checkpoint).

Results

Neither regime reliably contracts an injected error

Open-loop. Across all three stamp types, unstable states are more common than stable states among measurements that clear the gates, while most stamps fall in the deadband. The deadband is not noise: 90–99% of stamps that clear Gate 1 also clear Gate 2 at \(K{=}24\). It is driven by rates too small to cross the amplification threshold; on lift, 78% of demonstration stamps have \(|\lambda| < 0.0289\). Within robomimic the median rate rises from can (\(0.009\)) to lift (\(0.015\)), square (\(0.017\)), and tool hang (\(0.026\)), the most precision-demanding task. Rollout medians stay close to their demonstration counterparts (lift \(0.019\) vs. \(0.015\)), so moving from expert states to policy-visited states does not substantially change how the plant propagates a small error.

Distributions of the measured propagation rate pooled by benchmark family, for demonstration, rollout, policy-chunk and closed-loop stamps
Distributions of the measured propagation rate \(\lambda_{\mathrm{mean}}\), pooled within each benchmark family. Columns: robomimic, MimicGen, RoboCasa; rows: demonstration, rollout, policy-chunk, and closed-loop stamps. Verticals mark the reference Gate-1 thresholds \(\pm\ln(m)/16\) for \(m\in\{1.5, 2.0, 3.0\}\). Demonstration, rollout, and closed-loop rates are fitted over \(K{=}24\); the policy-chunk rate is capped by the emitted chunk length.

The error is front-loaded. The perturbed branches separate rapidly in the first few steps and then flatten rather than keep growing at the same rate. The measured rate therefore depends on the fitting window: a short window fits the initial rise and reports more unstable states, a long window averages that rise with the plateau and pulls \(\lambda\) toward zero. \(\lambda\) is a summary of propagation over an interval, not a time-invariant rate for the whole trajectory, and a short chunk can end while the error is still growing where a longer one continues through the plateau.

Closed-loop. Closed-loop stability is substantially rarer than passive contraction under the same measurement: at \(m{=}2\) the stable class holds only 3–88 stamps per task, despite hundreds of non-stable stamps. Comparing the two regimes at the same policy-replan states, the [OL unstable, CL stable] cell is near zero on all three benchmark families. States that amplify an error in open loop rarely contract it when the policy replans; they move to the deadband or remain unstable. Replanning trades away the temporal commitment of a chunk for reactivity without reliably reducing error propagation.

This exposes a gap in standard imitation training. The policies are trained directly from expert demonstrations, without perturbation augmentation or recovery supervision, which constrains them on the demonstrated states and says little about recovery after a deviation. Closed-loop contraction may need to be encouraged explicitly by training on perturbed states and their recovery actions, in the spirit of DART (Laskey et al., 2017) and Simchowitz et al. (2025).

Prediction

Can the regime be read from what the robot observes?

The open-loop rate is a property of the state the robot is in, so we ask whether it is visible. A frozen V-JEPA 2 ViT-L encoder over a short window of camera frames, with proprioception over the same window, feeds a two-layer head that predicts \(\lambda\) and a logit for the unstable class. P receives frames and proprioception only; A1 additionally receives the action chunk under evaluation. Classification uses only stamps that clear both gates; regression uses all stamps.

Open-loop is visible. On demonstration and rollout stamps the predictor reaches AUROC between \(0.63\) and \(0.94\) where the negative class is sufficiently supported, strongest on tool_hang and mg-coffee. Regression is weaker, recovering a tenth to a quarter of the constant-median error: the model is better at identifying which states are prone to amplification than at predicting how much the error will amplify. Adding the action sequence improves regression and rank correlation, as expected when the target depends on the sequence replayed.

Closed-loop is only partly recoverable. With the same predictor, inputs, and stamps, switching the target to the closed-loop rate still improves on the constant-median baseline on most regression cells, and AUROC exceeds \(0.70\) in half of the cells with both classes present, but the gains are less uniform. The closed-loop rate depends not only on the configuration but on how the policy acts after the perturbation.

DatasetInput Open-loop targetClosed-loop target
MAE0MAE\(\rho\) MAE0MAE\(\rho\)
liftP.0792.0535.757.0322.0178.819
liftA1.0792.0400.795.0322.0164.841
canP.0981.0635.706.0395.0323.520
canA1.0981.0568.762.0395.0305.553
squareP.0877.0790.510.0370.0364.332
squareA1.0877.0742.560.0370.0357.386
tool_hangP.0887.0687.627.0377.0393.239
tool_hangA1.0887.0664.637.0377.0347.333
mg-coffeeP.0833.0842.149.0400.0311.436
mg-coffeeA1.0833.0835.215.0400.0317.455
mg-nutP.0959.0941.284.0391.0373.374
mg-nutA1.0959.0927.324.0391.0344.474

Prediction of open- and closed-loop propagation rates at the same policy-replan stamps. Both heads use the same trunk and inputs; only the target differs. \(\mathrm{MAE}_0\) is the constant-median baseline, \(\rho\) is Spearman correlation with the measured rate. Bold denotes improvement over the baseline.

Conclusion

What state-wise measurement changes about the story

We studied what happens after an action error enters a chunked imitation-learning system, rather than how often the policy introduces one. Confidently stable states are uncommon, amplification is common among states whose rate is resolved, and replanning frequently removes confident amplification but rarely produces confident contraction. Propagation is strongly state- and horizon-dependent: even within a single task, rates vary substantially across states, so the assumption of a single global contraction rate is not supported empirically even at the task level, let alone across a benchmark.

These results do not determine the best execution horizon by themselves; stability is one factor alongside task performance, temporal consistency, and replanning cost. They do suggest that closed-loop error recovery should not be expected to emerge reliably from standard imitation learning. If it is desired, it should be trained explicitly, by exposing the policy to perturbations and recovery trajectories through perturbation- and tree-coverage-oriented training.

Label videos

What the labels look like

Each clip starts at one labeled state: the moment we nudge the first action. From there it plays the next \(K{=}24\) time steps, which is the whole window the label is measured over, so the clips are short by design. The white path is the robot's end-effector when nothing is nudged; the coloured paths are the \(N{=}8\) nudged copies. Watch whether the coloured paths pull back toward the white one (stable), fan out away from it (unstable), or neither (deadband). The label and its rate are printed in the corner.

nominal end-effector path perturbed, unstable perturbed, deadband perturbed, stable

Demonstration stamps, \(K{=}24\), \(m{=}2.0\). Buttons 1–2: unstable, highest \(\bar{\lambda}\) clearing both gates. 3–4: deadband, \(\bar{\lambda}\) nearest zero. 5–6: stable, most negative \(\bar{\lambda}\) clearing both gates. Hover a button for the stamp.

BibTeX

@misc{goyal2026stability,
  title         = {Measuring the Stability Assumption Behind Action Chunking},
  author        = {Goyal, Aryan},
  year          = {2026},
  eprint        = {2610.01626},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.01626},
}