acceptodds
Withdrawn from ICLR 2027

Speed on Demand: Faster-than-Demonstration Imitation via a Neural Robot Dynamics Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Imitation policies trained on teleoperated demonstrations succeed on a broad range of manipulation tasks, but their execution speed is typically capped by the frequency at which human demonstrations are collected. Naively raising the execution frequency at deployment fails, as the tracking error between commanded and reached joint positions grows with frequency, so the executed trajectory deviates from the demonstration and the policy's observations fall out of distribution. We present Speed on Demand, a framework that enables a BC policy to operate faster than the demonstration by learning intentionally offset commands whose reached positions match the demonstrated trajectory after passing through the robot–controller dynamics. A neural dynamics model, trained using only 2 minutes of play data, captures the command-to-reached-state mapping of the robot–controller system. By fine-tuning the BC policy with this neural dynamics model in the loop, the policy learns to produce intentionally compensated commands whose resulting reached positions match the demonstrated trajectory under accelerated execution. A single policy, conditioned on a continuous speedup factor, executes at up to the demonstration frequency on CRANE-X7 with 80% task success, where both baselines fall to 27% or below. On Franka Research 3 with impedance control, the same method reaches the demonstration frequency, and task success is preserved up to . Videos and code are available on the https://iclr90866-anonymous.github.io/project website.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Imitation Learning, Demonstration Acceleration, Neural Dynamics Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_791TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Summary

This paper proposes Speed on Demand, which accelerates imitation policies by compensating for frequency-dependent tracking errors. A neural dynamics model trained on play data predicts reached joint positions. The pretrained policy is fine-tuned through this frozen model to match demonstrated positions rather than commands. A speedup factor conditions the policy for execution at different frequencies. Experiments cover three manipulation tasks on two robot platforms. The method achieves 80% success at five times the demonstration frequency on CRANE-X7. Results on FR3 are mixed: mug hanging improves over direct acceleration, while high-frequency cup insertion suffers from overcompensation and reduced task success rates.

Critical strengths

Strengths

  • The central idea is straightforward: use a learned dynamics model to directly fine-tune a pretrained policy to compensate for frequency-dependent tracking errors.
Critical weaknesses

Weaknesses

1. The added value and scope of the methodological contribution need to be better established.

My understanding is that the main methodological novelty lies in using a learned dynamics model to directly fine-tune the policy for different execution frequencies. Compared with SAIL's execution-side mechanisms, this provides a conceptually more straightforward approach. However, this simplification introduces a dependence on the dynamics model's accuracy, particularly for the compensating commands generated during policy fine-tuning. I am concerned that model mismatch could substantially restrict the method's applicability.

2. Important baseline comparisons are missing.

  • Methods with access to online interaction data. The proposed method collects additional robot interaction data to train its dynamics model. It should therefore also be compared with methods that leverage online interaction, rather than only direct acceleration and gain adjustment. For example, SpeedTuning uses RL to learn an execution-speed policy on top of a base manipulation policy. What advantages does the proposed approach offer over such alternatives under comparable interaction budgets, accounting for differences in supervision requirements?

  • A representative comparison with SAIL. SAIL is not simply a high-gain controller baseline. If the DiT High-Gain's base policy is only outputting the commanded pos and force the robot to stay closer to the commanded pos, the failure is obvious, which is also stated in SAIL's appendix. Comparing only against DiT High-Gain does not sufficiently establish the benefit of dynamics-model-based policy fine-tuning over SAIL, which the authors identify as the closest related method. A direct comparison is needed to clarify their respective benefits and limitations.

  • A simple baseline is missing to assess whether the forward dynamics model is necessary. The current baselines—direct acceleration and increased controller gain—do not sufficiently isolate the benefit of the proposed forward-model-based policy fine-tuning. A straightforward alternative would be a two-stage approach: first train a policy to predict desired reached positions, then freeze this policy and train a separate frequency-conditioned module to convert these targets, together with recent joint-state history, into commanded positions. This conversion module could be trained directly on the same multi-frequency interaction data.

    • Although this alternative would still learn an inverse mapping, it would not require a separate forward dynamics model or backpropagation through that model to update the task policy. Comparing these approaches under the same data budget would help determine whether the proposed fine-tuning mechanism provides additional value over separating task prediction from execution compensation. Without this comparison, it remains unclear whether the reported improvements require the proposed forward-model-based approach.
Questions for authors

See weaknesses

Reviewer_bLVjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a method to execute BC policies at faster than demonstration frequencies. The authors train a small dynamics model on play data, which is replayed at several frequencies and train a BC policy with flow matching, with a smooth trajectory bias on the demonstrations. Then they finetune the BC policy: conditioned on the desired frequency, the model is finetuned on mean squared error of the outputs of the dynamics model and and the final reached state. On Crane X-7 and FR3 tasks, the authors show that this method keeps task success rates high throughout a various number of sped up frequencies.

Critical strengths

The paper clearly shows that the problem of executing faster than demonstration BC policies by naively increasing the execution frequency is real in their experiments and that their method is cheap to train, requiring only 2 min of additional play data.

Critical weaknesses

My biggest concern with the paper is that it does not justify the reason for proposing their method clearly. For example, why can't we train a BC policy on subsampled steps, conditioned on the execution frequency? If the demonstrations were given at 50Hz and the desired execution frequency is 100Hz, can you not train the BC policy on (s_t, f=100/50) -> a_t \dot a_{t+1} where \dot denotes the composition of actions? Or one could even try an inverse dynamics model (s_t, s_{t+k}) -> a_{t+k} and train the BC policy directly on those actions. The paper does not seem to explore these other ideas or discuss why these would not be useful.

Questions for authors

Comments and questions:

  • A_t seems overloaded in Eq. (3), means both commanded joint target and also the compensated command
  • Eq. 6 seems to be the standard flow matching objective, how does TT-RTC fit in?
  • There doesn't seem to be any baselines, are there any other published papers to compare with?
Reviewer_ZnjZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies faster-than-demonstration execution of behavior cloning policies in robotic manipulation tasks. Authors argue that naively raising the control frequency to speed up execution causes reached joint positions to lag behind the commanded targets.

Their proposed method trains a small MLP that predicts the next reached joint position from robot arm's recent three steps joint history, the commanded target, and the control frequency, using about two minutes of free-space play data replayed at up to 3× the demonstration's control frequency.

A BC DiT policy is first trained via flow matching loss on the demonstrations. Then this policy is finetuned conditioned on a continuous speedup factor, with a loss that passes its commands through the frozen dynamics model and penalizes the gap between the predicted reached positions and the demonstrated ones.

Authors evaluated this method on three real-robot tasks (CRANE-X7 object sorting, FR3 cup insertion and mug hanging) with 15 rollouts per frequency rate against naive acceleration and a higher-gain controller on the CRANE-X7 arm. It reports 80% success at 5× on the CRANE-X7, but on FR3 cup insertion beyond 2.5× its success falls below naive acceleration.

Critical strengths

There is very limited strength to this paper. The only one I could think of would be:

  1. This paper experimented proposed method on two different robot arms with different controller types: position controlled CRANE-x7 and impedance controlled Franka arm.
Critical weaknesses
  1. There is no comparison to prior acceleration methods or to simpler compensation baselines. The paper cites SAIL, DemoSpeedup, SpeedTuning, RACE, B-spline Policy, and I2RLC. It compares only against naive acceleration of its own BC policy and a P-gain increase described as one component of SAIL conducted on the CRANE-X7 only. That gain baseline already drops to 73% at 1× (vs. 100% at the default gain), so it degrades the policy before any speedup and is not a meaningful replacement for SAIL.

Simpler approaches to address the same tracking lag are also untested: a frequency-tuned command lead, analytic feedforward compensation from a fitted tracking model, inference-time inversion of the learned dynamics model without policy finetuning, or offline relabeling of demonstrated actions with model-inverted commands then fitted using standard flow-matching training.”

The task setup is thinly described: initial object positions vary by about 2 cm, while target / goal variations, eval protocols, and success criteria are not specified. It is therefore unclear whether the tasks require closed-loop policy behavior at all, as opposed to time-scaled replay with compensation.

  1. The design space is not ablated, and the mechanism behind the performance gains is not experimentally proven. The finetuning procedure is underspecified. The paper does not state whether chunks come from full multi-step sampling or a one-step estimate, how many integration steps are differentiated through, or whether the flow-matching loss is retained, even though reward-gradient fine-tuning of flow models is known to need truncation or a principled formulation (e.g., DRaFT, Adjoint Matching).

The dynamics model is also under examined: its prediction accuracy is never reported, it is trained only up to 3× but used up to 5×, and its three-step history, which the authors blame for the FR3 cup-insertion collapse, is never varied.

Questions for authors
  1. Finetuning procedure: How is the action chunk produced during fine-tuning: full multi-step sampling and integration from Gaussian noise, or a one-step estimate from a noised demonstration chunk?

How many integration steps are differentiated through, and is the flow-matching loss (Eq. 6) retained? If L_reach is the sole objective, does the fine-tuned policy preserve sample diversity?

  1. Dynamics model accuracy: What is the prediction error of at each frequency? Does this error correlate with the task-success drops, particularly on FR3 cup insertion?

  2. Speedup measurement: How is the achieved speedup computed? Why does it exceed the commanded speedup on mug hanging (3.6× at 2.5×)? Does the fine-tuned policy shorten the trajectory or skip pauses, rather than time-scale it?

  3. Task setup: Do target locations (bin, receiving cup, mug stand) vary across episodes? What does the policy observe, and how is success criteria determined? Would time-scaled replay of the nearest demonstration succeed under this 2 cm variation?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.