- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a coarse-to-fine approach for pretrained flow-matching VLA policies by modifying the sampled Gaussian noise. The method constructs an orthonormal basis for coarse trajectory commands and trains the flow-matching model to generate residual fine-grained motion. This separates where to go from how to move. The authors evaluate the method on quadrotor gate navigation and out-of-distribution maneuvers in simulation and the real world.
Critical strengths
-
Interesting mathematical formulation. The method uses the flow-matching interpolation to preserve coarse trajectory commands rather than relying on additional conditioning or inference-time guidance.
-
Clear motivation. The separation of where to go and how to move is intuitive, and the mathematical formulation is easy to follow.
-
Real-world evaluation. The authors evaluate the policy on a physical quadrotor, including unseen orbit and figure-eight trajectories. They also use Gaussian Splatting to reconstruct the environment for simulation and evaluate robustness under visual distribution shift.
Critical weaknesses
-
Limited evaluation trials. In Figure 7, the real-world OOD experiments include only five trials per task and lack baseline comparisons. The simulation experiments also use only ten trials for orbit and figure-eight tasks. More trials and real-world baseline comparisons would strengthen the results.
-
Choice of coarse motion basis. The authors manually construct the 16-dimensional command space. A comparison with PCA-based representations extracted from demonstrations would help validate this design choice.
-
Limited task diversity. The evaluation mainly focuses on quadrotor gate navigation. Additional tasks with different motion requirements would make the generalization claims more convincing.
Questions for authors
-
Missing ablation on command prediction. In Section 5.2, the authors claim that low-dimensional coarse motion is easier to predict under visual distribution shift. It would help to compare against a baseline that predicts the full action chunk with a comparable prediction head to verify this claim.
-
Additional steering baselines. It would strengthen the paper to compare against other steering approaches, such as residual flow steering or learned noise-space steering methods.
2 comments
Response to Reviewer KrLh (Part 2/2)
W3: Limited task diversity
The evaluation mainly focuses on quadrotor gate navigation. Additional tasks with different motion requirements would make the generalization claims more convincing.
We agree that the evaluation centers on quadrotor gate navigation, and we will state this as a limitation. We note that our construction needs only a linear map from action chunks to a low-dimensional subspace, so other embodiments may use the same training objective with a customized . Within the drone setting, we acknowledge that our tasks are all navigation and trajectory-following tasks, though each captures a different facet of navigation:
- Gate navigation requires precision, as the drone must pass through a narrow opening without touching the posts; the arbitrary gates test this at unseen gate placements.
- The orbit and the figure-eight require dynamic movement absent from the training data, with continuous turning and repeated changes in direction.
- The mannequin tasks add a semantic search component to the task as its position is not known in advance.
Q1: Missing ablations on command prediction
In Section 5.2, the authors claim that low-dimensional coarse motion is easier to predict under visual distribution shift. It would help to compare against a baseline that predicts the full action chunk with a comparable prediction head to verify this claim.
To verify this, we trained a baseline under the same conditions with the same head architecture to predict the full action chunk. Evaluated on 6,743 frames of real demonstrations, the 16-dimensional command is predicted with , compared to for the full action chunk (Table R5).
Table R5. Prediction of each head against its own target on the real demonstrations.
| Head | Target | (95% interval) |
|---|---|---|
| Ours | 16-dimensional command | 0.39 (0.36, 0.42) |
| Full chunk | action chunk | 0.23 (0.20, 0.25) |
We also note that the baseline predicts the full action chunk on the same backbone, using the flow-matching action expert in place of a head. Relative to the baseline, the 16-dimensional command gets the drone through the gate opening more often, even when . In the runs of Figure 6, our policy misses the opening in 44/200 flights (a breakdown we will add to the paper) whereas the baseline misses the opening 69 times (Section 5.2). Most of our policy’s remaining failures are touches of a post (Section 5.3), which the learned then corrects, raising success from 87/200 to 146/200, compared to 89/200 for the baseline.
Q2: Additional steering baselines
It would strengthen the paper to compare against other steering approaches, such as residual flow steering or learned noise-space steering methods.
We see methods like residual flow steering and learned noise-space steering as complementary to ours, since they address a different question. Learned steering methods such as latent-space reinforcement learning (Wagenmaker et al., 2025) and flow reversal steering (Tang et al., 2026) adapt a trained policy toward an objective through additional optimization or interaction, so a fair comparison would require a reward and a separate training setup. Our method changes what the flow is trained to generate so that the coarse motion of an action chunk becomes an input that some entity (e.g., a learned head, a person, or an agent) can author at test time without reward or further training. The two can be combined, for example by running noise-space reinforcement learning directly in our 16-dimensional command space. For out-of-distribution route following (Section 5.4), our baselines are training-free methods of imposing one command on the fine-tune through manipulation of the source noise. We will add this discussion to the related work.
Response to Reviewer KrLh (Part 1/2)
We thank Reviewer KrLh for their reading and constructive feedback. We are glad the reviewer enjoyed the mathematical formulation and motivation of our work. We address each concern and question below.
W1: Limited evaluation trials
We agree with the reviewer and have expanded the simulated orbit and figure-eight cells of Table 1 to 50 flights per method (Table R3).
Table R3. Orbit and figure-eight routes of Table 1 (50 flights per method). Err.: median tracking error.
Our policy completes 50/50 orbits and 50/50 figure-eights, whereas the π0 fine-tune with injected sketch completes 7/50 and 3/50, respectively. SDEdit (37/50, 47/50) and velocity projection (50/50, 50/50) maintain high success rates as well, consistent with Table 1. On hardware, we flew only our policy on the sketched routes for practical reasons: in simulation, the π0 baseline with the same sketch injected touches a gate in 43/50 orbits and 47/50 figure-eights, with median tracking errors of 44 to 74 cm. Flights that far off a route through gates risk catastrophic collision, which we deemed too costly to attempt. We agree that five hardware flights per route is a small sample. We will be sure to state that the hardware results demonstrate feasibility (that our policy can fly these routes) and are more of a robustness test than a contribution evaluation.
W2: Choice of coarse motion basis
We thank the reviewer for pointing this out, and we add a comparison of our command space against PCA bases (Table R4). We find that a 16-dimensional PCA basis captures more of the demonstrated motion and reproduces smooth translations more closely, but it cannot express motions rarely present in the training data. For example, commanding a 45 degree turn in place is off by about 34 degrees compared with 7 degrees for the windowed basis. We chose a basis that does not depend on the demonstrations because we want the command space to cover motions outside the training data. We would also like to note that the choice of command space is an interesting question in its own right, and we leave a systematic study as future work. We will add these comparisons in a revision.
Table R4. Expressing simple motions on the real demonstrations. We write each motion as a command, decode the command back into a path, and report the largest deviation from the intended motion (median over directions and over completing the motion in 20, 35 or 50 steps).