# Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation?

**ACT policies degrade at 2–3x the rate of their expert teachers when task execution speed increases — a finding that challenges the industry's standard success-rate-at-nominal-speed evaluation protocol.**

A new paper from Clinton Enwerem, John S. Baras, and Calin Belta, published today on arXiv (2609.01453), delivers one of the more uncomfortable findings for teams deploying [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) on humanoid hands: equal performance at nominal speed tells you almost nothing about how a learned policy will behave when execution tempo shifts. The researchers benchmarked a scripted expert against an Action Chunking with Transformers (ACT) policy on a contact-rich parcel manipulation task called ParcelStow — acquire, reorient, insert — and found that both achieve 100% task success at nominal speed. That equivalence evaporates under speed variation. At the maximum demonstrated speedup, expert success falls to 84%, while ACT success falls to 53%. Two independently initialized ACT policies show success-rate drops of 34 and 48 percentage points from nominal to maximum demonstrated speed; the expert drops only 16 points. The failure mode is specific: 35 of ACT's 47 failures at maximum speed are insertion misalignments, implicating fine contact control at the insertion phase rather than gross motion.

The results carry direct implications for anyone training humanoid manipulation stacks on teleoperated demonstrations, where execution speed is rarely controlled as a systematic variable.

---

## The ParcelStow Benchmark: What Was Actually Tested

The evaluation task — ParcelStow — requires the robot to acquire a parcel, reorient it, and insert it into a target location. It is explicitly contact-rich, meaning it exercises the kind of constrained, force-sensitive [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) that separates capable humanoid hands from glorified pick-and-place mechanisms.

The authors held task conditions and initial-condition draws constant across expert and learner, varying only speedup factor. Demonstrations span a defined speedup range applied to the manipulation phases after parcel acquisition. This is methodologically important: by fixing everything except speed, the experiment isolates temporal generalization as the variable of interest rather than conflating it with object or scene variation, which is where most robustness evaluations concentrate.

Both the scripted expert and the ACT policy trained on that expert's demonstrations achieve 100% at nominal speed — a result that would satisfy most deployment checklists in production robotics today. The divergence begins within the demonstrated speedup range, not beyond it. That nuance matters: the ACT policy is not being asked to generalize outside its training distribution in terms of speed. It is failing *within* the range it was shown.

---

## Where ACT Fails: Insertion Misalignment at High Speed

Stage-level analysis is where this paper becomes actionable. Of ACT's 47 failures at maximum demonstrated speed, 35 are insertion misalignments. The gross motion phases — acquisition and reorientation — hold up better.

The authors demonstrate this with a relative-motion handoff analysis: under that handoff condition, every ACT acquisition retains the parcel through reorientation and transfer in free space, but only 64% complete the overall task. After expert acquisition under the same conditions, 95% complete the task. The bottleneck is the insertion step, where accumulated timing errors in a contact-rich environment compound into misalignment that the policy cannot recover from.

This failure pattern is consistent with what the broader community has observed about ACT's sensitivity to temporal context. Action Chunking with Transformers was designed to handle multi-step action sequences, but the chunking horizon implicitly assumes a certain execution tempo. When that tempo shifts, the temporal alignment between predicted action chunks and actual robot state degrades — an effect that is more punishing at high contact forces than in free space.

From a hardware perspective, this is also a signal about actuator demands. Insertion tasks under speed variation stress both position accuracy and force regulation simultaneously. Whether a robot runs [backdrivable](https://humanoidintel.ai/glossary/backdrivability) quasi-direct-drive actuators or stiffer harmonic drives will shape how much the physical system compounds the policy's timing errors — something this paper's sim-based evaluation does not directly address but practitioners will need to account for.

---

## The Hidden Evaluation Gap

The paper's most pointed contribution is methodological: **nominal task success is not a proxy for temporal robustness**. This is not a subtle academic distinction. Teams at companies deploying [Physical AI](https://humanoidintel.ai/glossary/physical-ai) on humanoid platforms routinely use pass/fail at a single nominal speed as the primary gate for policy readiness. This paper shows that gate is insufficient.

The finding that none of the 414 acquisitions without force closure completes the task is a hard constraint that cuts across all evaluated policies and speeds. Force closure at acquisition is non-negotiable for task completion in this setting — a result that reinforces the importance of grasp quality metrics as a leading indicator rather than an afterthought.

For teams working with VLA architectures or [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) claims, the temporal robustness gap identified here is particularly relevant. If a policy trained by imitation degrades this sharply within its demonstrated speed range, extrapolating that policy to novel execution contexts — which is exactly what zero-shot deployment requires — introduces compounding risk that nominal-speed benchmarks will not catch.

---

## Industry Implications

The practical takeaway for humanoid manipulation programs is a two-part recommendation implied by the results, even if the authors frame it as an evaluation gap rather than a prescription:

**First**, add execution speed as a systematic dimension in policy evaluation. The speedup range used in demonstrations should map to the speedup range tested in evaluation, and success rate should be reported across that range, not just at a single nominal point.

**Second**, stage-level failure analysis — not just end-to-end task success — is necessary to identify where policies degrade. An aggregate 53% success rate at high speed obscures the fact that acquisition is nearly intact while insertion is catastrophically degraded. Those are different engineering problems with different solutions.

For companies like [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics) or [Figure AI](https://humanoidintel.ai/companies/figure-ai) pushing toward higher-throughput manipulation in logistics settings, execution speed is not an experimental variable — it is a core operational requirement. A policy that achieves 100% at demo speed but 53% at higher speed is not deployment-ready for any throughput-sensitive application, regardless of what the nominal benchmark says.

The code, data, and evaluation scripts are publicly available at the link in the paper, which should allow replication across different robot platforms and task configurations.

---

## Key Takeaways

- At nominal speed, both the scripted expert and ACT achieve 100% task success — the standard evaluation metric shows no gap.
- At maximum demonstrated speedup, expert success is 84%; ACT success is 53% — a 31-point divergence that standard evals would never surface.
- Two ACT policies with different parameter initializations drop 34 and 48 percentage points from nominal to maximum speed; the expert drops 16.
- 35 of 47 ACT failures at maximum demonstrated speed are insertion misalignments, not acquisition or reorientation failures.
- None of the 414 acquisitions without force closure completed the task, across all policies and speeds.
- Nominal task success does not imply preservation of expert performance across execution speeds — the paper's core claim, supported by controlled experimental data.

---

## Frequently Asked Questions

**What is temporal robustness in robot manipulation?**
Temporal robustness refers to a policy's ability to maintain task success as execution speed varies. A temporally robust policy succeeds not just at the speed it was trained on but across a range of faster or slower execution tempos.

**What is Action Chunking with Transformers (ACT)?**
ACT is an imitation learning architecture that predicts multi-step action sequences (chunks) rather than single actions, allowing robots to execute smooth, coordinated motions learned from human demonstrations. It is widely used in dexterous manipulation research.

**Why does imitation learning struggle with speed variation?**
Policies learned by imitation capture the temporal structure of demonstrations at a specific execution speed. When speed changes, the alignment between the policy's internal timing assumptions and the actual robot state degrades — particularly in contact-rich phases where small timing errors produce large positional errors.

**What is the ParcelStow task?**
ParcelStow is the contact-rich manipulation benchmark used in this paper, requiring a robot to acquire, reorient, and insert a parcel. It was designed to stress fine contact control, making it sensitive to the kind of temporal misalignment that speed variation introduces.

**What should robotics teams change about their evaluation protocols based on this research?**
Teams should evaluate manipulation policies across a range of execution speeds, not just at a single nominal speed. Stage-level success rates — broken down by subtask — provide more diagnostic signal than end-to-end pass/fail rates, particularly for identifying whether degradation is in gross motion or fine contact control phases.