- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This work is about guarantees for learned kinematic maps (for instance, mapping a robot arm position to the kinematic angles of its joints). The guarantees center around topology, conditioning, and statistics of the dataset.
- On topology, the standard is to measure with , minimum number of continuous joints needed to invert the map and , the minimum number for path planning. The paper shows that it is sufficient for the model's prediction to be within some error threshold everywhere, meaning that if if , there must be a large error somewhere in the network. If the network is K-Lipschitz, that large error is computable and we can get a lower bound on MSE and the necessary size of validation dataset. On the circle they get average error /.
- On conditioning, they consider the case of the single-joint robot arm (2R arm) where B_0=1 but the inverse blows up like . Author show how this steepness shows up in the sampled budget with a regression bound.
- On statistics, they consider the case where topology does not determine the sample complexity (). So the rest is link-length calibration. Experiments are on the 2R and 7R arms. Numerical IK outperforms the learned models and authors demonstrate their proposed bounds.
Critical strengths
- The paper proposes a practical set of findings in challenges in learning kinematic maps on topology, conditioning, and statistics. This conceptual framework can reasonably inform choices in evaluation and model design. On the evaluation side, there appears to be concrete value in reporting worst-case metrics rather than just RMSE (Thm 4 shows that on the circle, avg error can keep decreasing while worst-case error stays constant). On the model side, results show a principled method () for determining whether a single continuous network is sufficient or if we need more heads/labels (see weakness below that distinguish this great theoretical motivation from actual experiments on this).
- The main derivations (circle frontier, Smith count of , calibration risk) are well written and convincingly computationally checked. Code and checkpoints are provided and reported metrics are reproducible.
Critical weaknesses
-
The experiments don't seem to test the main topological theoretical findings. Every learned model (for 2R and 7R arms) is trained on a map, and the circle results use an analytic extremizer (not a trained network). So this means Thms 3-4 aren't checked with a network learning a (non-trivially) topological map.
-
I am seriously concerned about the paper’s overall clarity and writing structure/style outside of the derivations. Many sentences make little sense or are so dense that they take a lot of effort to parse. There are also a lot of AI-style “disclaimers” in almost every paragraph, which make the main message hard to follow.
Some examples of very dense or non-sensical sentences (not a complete list):
- “Hence, coordinate reconstruction alone does not solve hyper-redundant RNA loop proteomics.” (l. 40–41). RNA isn’t proteomics?
- “Algebraic topologists espoused their views on complexity in terms of invariants Pavešić (2019); Zapata & González (2023).” (l. 61) I have no idea what this means.
- “Planar and spatial seven-joint studies, including sixty retained spatial regressors and independent Jacobian checks, isolate representation effects and redundant-solution ambiguity.” (abstract, l. 23–24). Is "sixty retained spatial regressors” just 60 trained networks?
- “Our contribution is to establish explicit quantitative connections and separating examples (from manifold reconstruction and computational hardness) — rather than new invariants or charts…” (l. 67–69). I think manifold reconstruction and computational hardness are only in the appendix?
Some examples of disclaimers (not a complete list):
- “This equality is definitional.” (l. 98)
- “The two classes overlap and do not exhaust constrained kinematic spaces. Homotopy dimension also need not equal regression dimension.” (l. 148–149)
- “It is not implied by either topological count.” (l. 242)
- “The input spaces and endpoint requirements explain the difference between the notions; changing a normalization by one does not remove that difference.” (l. 746–747)
- “A single known gauge does not resolve a noninjective observable.” (l. 1393–1394)
I am particularly concerned about this because the AI use statement reports only “concise editing,” although the writing suggests much heavier use of AI in drafting the manuscript. Even if the disclosure were updated, I am not sure the paper as currently written is easy enough to read and understand to be a valuable contribution to the field, or that it meets the standards of ICLR.
Questions for authors
Suggestions for future work
Can you train a network on a map with ? Currently the experiments don't consider a learned model with non-trivial topology.
Clarifications
Theorem 3 assumes a compact workspace, but the 2R and the 7R cases are open regions. So what happens to if we do restrict to a compact sub-workspace? Which experiments does theorem 3 apply to?
Could the extent of LLM use in drafting the text be clarified, to better clarify what “concise editing” means as currently reported AI use statement?