Paper: force-vs-pose decomposition, data-matched ALL, and leave-one-robot-out#10
Merged
5 commits merged intoJul 14, 2026
Merged
Paper: force-vs-pose decomposition, data-matched ALL, and leave-one-robot-out#105 commits merged into
5 commits merged into
Conversation
…orce-is-sensed) Q1/Q5 (sec:problem): force is a directly sensed instantaneous state variable (wrist F/T sensor), not differentiated from position; velocity/acceleration are true time-derivatives, out of scope for the single-timestep model. Q4 (sec:ablations): new 'Force isolated from TCP pose' study + table tab:forceonly. Force-only R2 0.05/~0/0.19 (flexiv/ur5/kuka); ee probe is pose-dominated (pose-only 0.41/0.31/0.53). Fused z_v beats raw ViT AND a compute-matched vision-only control on all three (5-seed paired t, p<=0.012); pose-partialled residual is ~0, i.e. recoverable force is pose-entangled at one timestep. Q2 (sec:ablations): new 'Data-budget-matched ALL' study + table tab:matched. Matched to a specialist's 32870-frame budget, ALL trails on-diagonal by 0.02-0.05 but still beats every specialist OFF its training robot -> embodiment coverage is diversity, not data volume. Reworded abstract, phase1 findings, and conclusion from 'recovers force ... substantially better' to the decomposed, honest version. Leave-one-robot-out (Q3) to follow in a second commit.
… embodiments sec:discussion: new paragraph + table tab:loro. Zero-shot LORO (train on 3 robots, probe the held-out 4th, 5-seed): the vision-only latent trails raw ViT on the held-out robot's motor probe by 0.03-0.13 (flexiv/ur5/kuka), so embodiment-agnosticism holds only for robots in the training mix, not zero-shot to unseen ones. Nuances: far more stable than raw on degenerate franka (-0.46 vs -6.71); ur5 force/ee zero-shot beats raw (+0.06). Framed as limitation + future work.
Ishneet0710
changed the title
It's Phase-2 (future-work) design and broke the flow of a single-timestep paper. Moved the subsection + its figure to a new appendix section 'Continuous-time, multi-rate extension (planned)', keeping the sec:temporal/fig:temporal labels so all 9 cross-references still resolve. The Method now covers only the single-timestep model actually evaluated.
Reframe around the vision-sufficiency null (H0). Experiments now read good->useful: a two-act §5 (A good representation: transfer matrix; A useful representation: downstream, safety-first) followed by the controls (ablations). Promote/rename downstream from a buried section to the useful-representation subsection. Declutter body to appendix: patch-feature PCA + attention maps (Interpretability probes), bottleneck/joint-SIGReg + triplet (Secondary ablations), the freeze-backbone preliminary, and the continuous-time extension (planned). Remove the franka-heavy transfer-matrix heatmap (tables kept). Tighten prose to bold-run-in, concise paragraphs (Findings, downstream, Discussion); condense Data to four paragraphs; light-trim Related Work and the Method rationale. Rewrite Abstract/Intro/Contributions/Conclusion to lead with the latent + downstream-safety; drop the evaluation-protocol contribution; add series-positioning (action predictor + native-rate temporal as next steps). Fix: define z_v in the problem formulation (was used before definition), typos (perceivable, end-effector), and a cite->citep. Compiles clean: 0 undefined refs/citations, 0 duplicate labels, 26pp.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds three experiments to the analysis and tightens the claims they bear on. Paper-only; 2 signed commits; compiles clean.
New experiments (5-seed):
Wording: