- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes an execution-level adapter for frozen vision-language-action (VLA) policies. A predictor calibrated on healthy motion estimates expected robot responses; an observer uses prediction residuals to estimate additive command offsets and applies bounded corrections. The paper also introduces crossed command-stream replay to distinguish correction effects from effects associated with changed policy commands.
On Panda/LIBERO, adaptation raises faulted success from 58/220 to 121/220 on held-out initial states evaluated under two sampler seeds. A separate unseen-offset evaluation improves success from 48/160 to 125/160. Additional simulations examine other policies, robot interfaces, and limits imposed by correction authority, delay, and predictor mismatch.
Critical strengths
-
Substantial recovery with credible held-out evidence - Faulted success improves by 28.6 percentage points [16.2, 41.5] in the main holdout, with similar gains under both sampler seeds. The separate unseen-vector evaluation further supports generalization beyond one offset. Healthy controls are included, and historical evaluations with selection overlap are clearly distinguished from holdout evidence.
-
A useful diagnostic protocol with appropriately cautious interpretation - Crossed replay imposes recorded command streams while allowing the observer to update from its own measurements. The reference-sensitivity analyses and task-level counterexamples strengthen the study by showing what the trajectory contrasts can and cannot establish.
Critical weaknesses
-
The advantage of the specific NT observer is unclear - Under equal candidate budgets, the nominal-error integral controller achieves 138/160 unseen-fault successes versus NT’s 125/160, with a paired advantage of 8.1 points [1.9, 14.4]. Healthy-performance differences remain unresolved. The APEX-style comparison does not establish superiority over published APEX because it is explicitly an approximate reconstruction. The results support calibrated execution feedback more strongly than the particular NT design.
-
Replay conclusions depend on the reference and retained population - All five tested π₀.₅ NT stream-effect intervals include zero under alternative healthy references. Moreover, matched faulted π₀.₅ replay retains 27/40 cases, while GR00T NT retains 30/40 and omits two entire tasks. These checks are transparently reported, but leave uncertain whether the physical effects characterize the full task-evaluation population.
-
Cross-robot coverage does not establish a common online recovery method - Panda adapts within the evaluated episode; ALOHA, GR1, and WidowX estimate corrections in preceding faulted episodes and then hold them fixed. GR1 and WidowX also use different healthy-control schedules. Broader physical-fault evidence is weaker: the combined friction-shift gain is unresolved, and locking remains unsuccessful.
-
The theory remains primarily explanatory - Response geometry characterizes realized trajectories, while the memory–authority–delay result assumes idealized scalar dynamics and exact fault-sign information after a specified delay. The prospective study produces no decisive benefit bounds, and its sign accuracy matches an always-benefit baseline. The analysis therefore does not yet offer a demonstrated rule for anticipating when adaptation will help or harm.
The technical contribution is primarily in execution-level adaptive control. While the interaction between correction and frozen VLA behavior is relevant to ICLR, the control analysis is more developed than the new insights into learned-policy behavior, which limits the paper’s contribution to the machine learning venue.
Questions for authors
-
What demonstrated advantage motivates choosing NT over the nominal-error integral controller?
-
How do retained and excluded replay cases differ in task outcomes and available trajectory characteristics?
-
Can the non-Panda results be evaluated under matched online and healthy-control protocols?
-
Can the analysis predict beneficial or harmful adaptation on new executions beyond a simple baseline?
2 comments
The APEX-style comparison
We agree with you; it compares NT with our own reconstruction and cannot show superiority over published APEX. We could not reproduce APEX's discrete coefficient updates from the published description, so the revision reports the reconstruction only in Appendix C.7. We also ran A2C2, a delay-correction head learned from demonstrations, with its authors' code (item 9, Appendix F.7).
Q2. Retained and excluded replay cases
Excluded cases are shorter and more often solved without correction, so the retained cases over-represent the policy's failures, and the physical effects describe only the retained cases. Cases drop out when a run is shorter than the 80 steps replay needs, when runs diverge before correction starts, or when replay fails to reproduce the recorded motion; Appendix F.9 gives the counts. Table entries are task successes without and with correction, and each row's retained and excluded cases add up to the 40 assigned to replay.
| Policy and rule | Retained cases | Excluded cases |
|---|---|---|
| π0.5, NT | 2/27 → 9/27 | 4/13 → 5/13 |
| π0.5, DOB | 2/27 → 8/27 | 4/13 → 6/13 |
| GR00T, NT | 0/30 → 8/30 | 4/10 → 3/10 |
| GR00T, DOB | 0/29 → 6/29 | 4/11 → 4/11 |
The median fault-free episode lasts 81 steps in excluded GR00T cases, and 106.5 in retained ones (92 and 107 for π0.5), and GR00T loses two whole tasks to the checks. The only two episodes that correction turned from success into failure, one per policy under NT, are both excluded, so the physical effects of all four replay sets come from cases that correction saved or left unchanged. Task success still counts all 40 cases, and the π0.5 episode in Section 5.2 that fails only with correction comes from an earlier replay collection.
Q3. Matched online and fault-free protocols beyond Panda
For ALOHA, yes. GR-1 and WidowX keep their original evaluations, which estimate the offset in preceding faulted episodes and hold it, without matched fault-free controls, and Section 6 now lists this as a limitation.
On 40 new ALOHA seeds, with fault-free controls on the same seeds, we ran adaptation from a zero estimate within each episode under a +0.02 rad offset on the six left-arm joint targets. We first ran NT with its submitted settings, then ran all conditions below together in a second run, which the table reports (successes out of 40).
| Correction | Fault-free | Under the offset |
|---|---|---|
| None | 14 | 0 |
| NT, submitted (gain 0.08) | 13 | 0 |
| NT, faster (gain 0.5) | 13 | 10 |
| DOB, faster (gain 0.5) | 15 | 18 |
| Estimated on one preceding episode, then held | 11 | 10 |
| True offset subtracted | n/a | 13 |
With the submitted settings, NT repairs none of the faulted episodes. An earlier observer study (Appendix B.4) had selected the faster settings on other seeds. For NT, they also changed the cap and normalizer. With them, NT raises faulted success from 0 to 10 of 40 (+25.0 percentage points, 95% CI [12.5, 40.0]) and DOB to 18 of 40, while fault-free success is 13 and 15 of 40, against 14 without correction. DOB exceeds NT on this one task and fault, but we cannot distinguish either rule from the held estimate or from subtracting the true offset (Appendix F.11, item 12).
Physical faults beyond command offsets
We agree that this evidence is weaker. Friction and locking depend on the motion and need not be equivalent to any command offset. Correction partly repairs a joint-3 friction increase present from the first step (from 0 to 13 of 40 episodes), and joint locking scores 0 of 20 with and without correction. The new time-varying offsets (item 3) are still command offsets. In a new test (item 10), we applied the weight of a 1.0 or 1.5 kg payload to the hand. Uncorrected success falls from 73 to 65 and 51 of 80 episodes. Early identification, which estimates an offset from the first 20 control steps and holds it, restores success to 76 and 74, against its 75 without the weight, because in early motion the weight acts like a command offset. Section 6 states these limits.
Q4. Predicting benefit or harm on new executions
Our prospective forecasts of physical benefit always predicted that adaptation would help, so their sign accuracy equaled that of always predicting benefit. Section 4 now states that its results inform the choice of mask and cap but do not forecast task outcomes. Lower path deviation also need not mean higher task success (finding 4). The analysis did predict some physical effects before the runs. We expected constraining the predictor to raise the median rotation-y estimate late in the episode to at least 70% of the offset, and it reached 81.5% on LIBERO-Spatial and 70.5% on LIBERO-10; a ten-step delay and a halved cap raised the physical cost in 18 and 14 of 19 test episodes.
Thank you for the careful review. We agree that the results support calibrated execution feedback more than the normalized-target (NT) update rule, so the revision makes the feedback interface, with interchangeable update rules, its first contribution. New experiments test how frozen policies respond to offsets and correction, and evaluate adaptation within the episode on ALOHA with matched fault-free controls. "Item N" refers to row N of the results table in our general response.
Insight into learned-policy behavior
The revision reports five findings about VLA policies, all in simulation. Findings 1, 2, and 5 are new, and the revised introduction states the first four.
Q1. Advantage of NT over the nominal-error integral
We found none, so the revision treats the update rule as a replaceable part of the interface, which carries the contribution. We ran all update rules together on the 220 held-out Panda evaluations. The nominal-error integral, which integrates the deviation of measured motion from the motion predicted for the policy's own commands, succeeds in 145, NT in 119, and no correction in 60 (integral minus NT +11.8 percentage points, 95% CI [4.7, 19.3]), in line with the result you cite on unseen offsets. Without the offset, every rule succeeds in 211 to 215 of 220, against 211 without correction (Appendix F.2, item 4).
After the runs, we found that the rules' settled rotation estimates follow the order of their stationary values from Section 4. NT attenuates the offset implied by each step's residual before averaging it, so its estimate settles furthest below the true offset, while the integral's settles near it.