- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
this paper presents PAP, which is basically an action primitive model for robot arm tasks combined with a learned local dynamics model. The dynamics model captures the system dynamics around a primitive action and predicts how the system responds when the interaction conditions change. With this capability, control can be computed based on model predictions, inverse dynamics, or rejection sampling. This idea is similar to model-based control methods like MPC, but the interesting part is that the paper proposes to learn a model to predict the system dynamics, which is different from previous work that trains a model to predict the next state. To achieve this, the paper first pretrains an encoder on a set of tens of millions of trajectories. This encoder can extract the necessary features from a few demonstrations, and the system dynamics prediction model is then trained on a couple of new demonstrations. The main paper does not provide comparisons with baseline methods (I am not sure if DMP and SPiRL are considered baseline methods, as the paper does not state so, nor does it discuss why they are chosen and what they are compared against). Based on my understanding, the system dynamics prediction model is quite overfitted to the specific tasks, and it can only handle small changes to the settings. The generalization and practicality of the proposed method are limited.
Critical strengths
the paper proposes to include a local system dynamics model as part of an action primitive, which is an interesting idea.
the paper proposes to train a model to predict the dynamics itself, which is a different approach from those that train a model to predict the next state.
Critical weaknesses
The main paper is not very well written. It appears to keep introducing new concepts and terminology (such as Interaction Laws, intervision, query, invoke, probe, etc.) whose meanings are not easy to guess, making it difficult to identify what is existing work can achieved and what the key technical contributions of the paper are. The key technical part, particularly how the system dynamics model is trained, is explained very vaguely and briefly. Even with the appendix, it is still hard to understand the details without guessing. Many metrics used in the paper, such as the audit discrepancy area in Fig. 4, do not appear to be commonly used metrics in the field, and the numbers are not obviously interpretable.
while I think the idea of learning a model to predict local system dynamics is interesting, I have strong concerns regarding the generalization that the proposed method can achieve. The three tasks used in the paper are simple and of similar types. The system dynamics are also very simple. The paper does not compare the method with more directly learned dynamics models, for example, as simple as a linear model trained using least squares with some intuitively selected features. On the other hand, if I understood correctly, the system dynamics model is kind of overfitted to the specific configuration of a task, which further limits its generalization ability. The paper may need to test on task with more complicated dynamics.
the paper lack of comparison with existing method, as I mentioned in summary
Questions for authors
please correct me if I am wrong with second paragraph in "Weakness"
8 comments
External baselines in the main paper, and the Fig. 4 metric in plain words
1. External baselines in the main paper (Fig. 4, Sec. 4.2). The main paper compares PAP against four external baselines—including DMP and SPiRL—in Fig. 4 and Sec. 4.2. As stated in the first paragraph of Sec. 4.2: "DMP (Ijspeert et al., 2013) represents measured tool trajectories, while SPiRL (Pertsch et al., 2021) represents state–action sequences … These comparisons test the physical information preserved by each organization criterion." The section then reports that "response-based neighborhoods have lower independently measured discrepancy than all four comparison organizations, with each paired interval below zero."
Why each baseline was chosen. Each baseline represents a widely used paradigm in robotics for deciding when two interactions belong to "the same reusable skill":
| Method in Fig. 4 | Paradigm it represents | Calls two pushes alike when |
|---|---|---|
| Oracle Semantic | symbolic action labels | they share the true verb label |
| DMP | trajectory primitives | their tool trajectories have the same shape |
| SPiRL | learned skill latents | their state–action sequences map to nearby skill latents |
| Endpoint | nominal outcomes | one nominal push gives the same measured outcome |
| PAP (ours) | local interaction laws | their predicted responses agree over a range of pushes |
2. Intuition behind the audit test and the Fig. 4 metric. Suppose an organizational method claims that two interactions are alike (neighbors). To test whether they are truly physically compatible, we apply 18 new, independent pushes to both objects and measure how their physical responses differ. In Sec. 4.2, we evaluate 320 query interactions against 640 stored gallery interactions on RollBall across five hidden damping values; each method locks its neighbor rankings before the 18 scoring pushes are collected.
- Plain-language meaning of the Fig. 4 numbers. The audit discrepancy is the measured difference between a query and its retrieved neighbors across the 18 new pushes, averaged over displacement, rotation, linear velocity, and angular velocity, each divided by its natural physical scale (5 mm, 5°, 0.05 m/s, 1 rad/s). A value of 0.25 corresponds to a 5 mm difference in displacement alone. The "area" simply averages this discrepancy across three neighborhood sizes (the closest 1%, 2%, and 5% of candidates; lower is better). PAP achieves 0.118, reducing discrepancy by 26.7% relative to Oracle Semantic (0.161) and by 19.1% relative to DMP (0.146), with all paired confidence intervals against SPiRL and Endpoint strictly below zero.
- Two directly interpretable percentages from the same recorded audit.
- Among pairs of pushes that share the same verb, the same initial context, and tool paths within 0.5 mm, 47.6% are physically distinguishable under new pushes (differing by more than 5 mm or 5° on at least 2 of the 18 test pushes). Neither a semantic label nor a trajectory representation can separate them.
- When we take pairs that start from the exact same state and differ only in hidden damping, and ask each method for the top 10% of pairs it considers most alike, the fraction of false matches (pairs whose measured responses are in fact physically different) is only 1.6% for PAP, compared to 4.7% for SPiRL, 26.6% for Endpoint, 43.8% for DMP, and 45.3% for Oracle Semantic (random chance: 52%).
Why controlled planar tasks are ideal for this test. Here, using clean, controlled tasks is an intentional design choice: it demonstrates that even in a simple setting without complex multi-stage contact, look-alike interactions conceal substantial physical response differences (47.6%) that standard semantic and trajectory primitives miss.
We will make the role of each external baseline explicit at the beginning of Sec. 4.2 and add the two intuitive percentages above.
Least squares, and how the model is trained
Thank you for suggesting this baseline. We implemented and evaluated two least-squares baselines under the exact Fig. 6 protocol. All four PAP constructions achieve lower prediction error than least squares fitted to the same seven pushes across all three tasks (12 of 12 comparisons), and our three nonlinear PAP constructions outperform both per-state least squares and feature-engineered least squares across every task.
Clarification on training: the model is trained once and frozen, never trained on new demonstrations.
- Each learned PAP construction is trained once on 900 training conditions pooled across the three tasks (300 per task, with no task identity label; Appendix B.5) and then permanently frozen.
- At test time, for a new state, the robot executes seven test pushes (probes). A single forward pass compresses them into a 32-number summary u (Sec. 3.2), with zero parameter updates.
- Every evaluation state is generated using a random seed unseen during training.
The two least-squares baselines we evaluated:
- Least squares on the seven pushes: an affine response model fitted directly via least squares to the seven test pushes at each state. Notably, this serves as the simplest linear instantiation of our PAP interface.
- Least squares with features: ridge regression fitted on the exact same 900 pooled training conditions as our neural models (16,200 held-out pushes), using intuitively selected features: a constant bias, the candidate push, the context h, the seven measured probe responses, and their cross-products with the candidate push. Regularization was tuned on the development set and locked prior to evaluation.
In the table below, prediction error is the sum of squared errors across the six response components, each divided by its physical unit (5 mm, 5 mm, 5°, 0.05 m/s, 0.05 m/s, 1 rad/s), where 1.0 corresponds to a one-unit error (e.g., 5 mm) in one component. Values are reported as PushT / PokeCube / RollBall.
| Model | Prediction error (lower is better) | Invoke success, % |
|---|---|---|
| Nearest of the seven pushes | 1.642 / 0.487 / 0.533 | 1.7 / 16.7 / 0.0 |
| Least squares on the seven pushes | 0.855 / 0.082 / 0.310 | 45.7 / 79.6 / 58.1 |
| Least squares with features | 0.595 / 0.054 / 0.203 | 39.8 on PushT |
| PAP-Affine | 0.618 / 0.060 / 0.213 | 38.8 / 80.2 / 99.6 |
| PAP-Nonlinear | 0.417 / 0.043 / 0.063 | 62.6 / 85.6 / 86.5 |
| PAP-Endpoint | 0.376 / 0.046 / 0.068 | 72.5 / 86.7 / 90.3 |
| PAP-ActionSeq | 0.395 / 0.037 / 0.054 | 72.5 / 87.0 / 100.0 |
- Response Prediction (Query). All four learned PAP constructions predict more accurately than least squares on the seven pushes in every task. Least squares with hand-crafted cross-product features closely matches
PAP-Affine. Crucially, all three nonlinear PAP constructions outperform both least-squares baselines across all three tasks (reducing prediction error from 0.595 to 0.376 on PushT, and from 0.203 to 0.054 on RollBall). - Intervention Selection (Invoke). On PushT and RollBall—where the physical response varies nonlinearly with the push—our nonlinear PAP constructions outperform least squares by 17–27 percentage points on PushT and 28–42 percentage points on RollBall.
We will add both least-squares baselines to Sec. 4.3 and revise Sec. 3.2 to make the one-time pooled training and frozen test-time protocol unmistakably clear.
Short clarifications, and how the paper fits together
Thank you very much for your thoughtful review, and especially for highlighting the core idea of our paper: proposing to "include a local system dynamics model as part of an action primitive", and training a model "to predict the dynamics itself" rather than merely predicting the next state.
Thank you also for inviting us to clarify the points in the second paragraph of your Weaknesses. We provide brief, direct clarifications below, followed by full experimental details in the following comments.
end-to-end Invoke comparison against external baselinesto Reviewer ZdaD).least squares, and how the model is trained).direct tests of overfitting: hidden physics).How our core contribution connects the three steps of the paper. As you perceptively noted, our approach predicts the local interaction law itself (how the object's physical response changes as a function of the candidate intervention), rather than predicting only a single next state. Our paper establishes this paradigm across three connected steps:
Importantly, the learned network is one concrete construction of a PAP; our formal definition of the primitive interface is model-class agnostic (Sec. 1)