Flow matching robot policies commonly use a fixed Euler step count selected from offline accuracy sweeps. At inference time, however, each policy input (visual observation, robot state, and language instruction) together with a source-noise draw defines a different ODE trajectory and sequence of local Jacobians. We ask whether a step count that accurately follows the nominal noise-to-action trajectory can still be too coarse for strongly contracting directions around it. An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged. We then study a frozen π0.5-LIBERO checkpoint without modifying its weights or action dynamics. Across 30 policy inputs from 10 LIBERO-10 tasks, the Jacobian exhibits strong contracting directions near the action endpoint, yet no sampled mode violates the Euler non-expansion condition. On five of those inputs, accuracy sweeps require 4--9 Euler steps at 5% terminal error and 87--112 at 1%. These five paired cases are therefore consistent with accuracy-limited rather than stability-limited sampling under this local criterion, while the controlled experiments show why endpoint accuracy alone cannot rule out a separate stability requirement.
Figures & tables
Figure 1: Accuracy and local stability impose distinct numerical requirements in Euler sampling. (a) Flow matching policy sampling from noise to action with N Euler steps. (b) Accuracy measures deviation from a refined reference trajectory, while local stability measures amplification of contracting Jacobian eigenmodes. (c) Either requirement can determine the necessary step budget. (d) For a frozen pretrained policy, we compare endpoint accuracy with a grid-aware local stability measure.
Regime
(β,κ)
Nacc∗
Nstab∗
N∗
Easy
(0.1,1)
1
1
1
Accuracy-limited
(10,1)
25
1
25
Stability-limited
(0.1,50)
1
25
25
Jointly limited
(10,50)
25
25
25
Table 1: Exact mechanism separation at ε=0.2 . Accuracy and stability independently set the admissible Euler budget, and the observed sweep matches the analytic thresholds in every regime.
Figure 2: Experiment 1: exact mechanism separation. Left: first-coordinate terminal error decreases with the Euler budget and is controlled by β . Middle: transverse amplification is controlled by κ and becomes non-expansive exactly at N=⌈κ/2⌉ . Right: the total terminal error combines both effects. Increasing β therefore changes the accuracy requirement, while increasing κ moves a separate, sharp Euler stability boundary.
Target
Base accuracy
Base stability
Full error only
Joint requirement
Eight-Gaussians
14.33±3.21
4.00±0.00
18.00±0.00
25.00±0.00
Bistable Langevin
21.67±0.58
1.67±0.58
21.67±0.58
25.00±0.00
Table 2: Controlled learned-flow augmentation at ε=0.05 , κ=50 (mean ± sample standard deviation over three seeds). The unaugmented checkpoints are accuracy-limited under the sampled local eigenmode test; the inserted mode then raises the joint requirement to 25 steps.
Figure 3: Experiment 2: controlled contraction added to trained flow matching models. Representative seed-0 checkpoints are shown for (a) Eight-Gaussians and (b) bistable Langevin. Left: the original learned-coordinate Euler error is unchanged across κ . Middle: the inserted transverse direction is amplified below its exact Euler boundary N=⌈κ/2⌉ . Right: the full augmented terminal error can become small before that transverse direction becomes non-expansive. Quantitative results over all three training seeds are reported in Table 2 .
Figure 4: Experiment 3: Pretrained-policy accuracy and stability results. Left: maximum sampled G=tχ← for 30 early-rollout policy inputs (three per task); G=1 marks the sufficient condition for local Euler non-expansion when all eigenvalues and times are covered. Right: persistent accuracy budgets for the five-input subset. The two panels have different coverage and should not be interpreted as a 30-input paired comparison.
LIBERO task
Nacc(5%)
Nacc(2%)
Nacc(1%)
Gmax
00
7
37
105
0.5115
02
9
39
112
0.5092
04
8
37
112
0.5142
06
4
27
87
0.5171
08
4
28
87
0.5088
Table 3: Paired accuracy and sampled-stability results on the five-input subset. Gmax is the largest value over the Arnoldi eigenmodes and prescribed reference times.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Target
κ
Nbase
Nstabbase
Ninsert
Npred=Njoint
Nfull
Rz−
Eight-Gaussians
0
14.33±3.21
4.00±0.00
0
14.33±3.21
14.33±3.21
–
10
14.33±3.21
4.00±0.00
5
14.33±3.21
14.33±3.21
5.06
25
14.33±3.21
4.00±0.00
13
14.67±2.89
14.33±3.21
2.61
50
14.33±3.21
4.00±0.00
25
25.00±0.00
18.00±0.00
6.83
Bistable Langevin
0
21.67±0.58
1.67±0.58
0
21.67±0.58
21.67±0.58
–
10
21.67±0.58
1.67±0.58
5
21.67±0.58
21.67±0.58
5.06
Appendix
Table 4: Complete Experiment 2 results at ε=0.05 . Nbase is the persistent accuracy threshold of the unaugmented learned field, Nstabbase is its sampled local-eigenmode stability budget, and Ninsert=⌈κ/2⌉ is the exact inserted stability floor.
Figure 5: Two-dimensional targets used in Experiment 2. Eight-Gaussians contains eight isotropic components arranged on a circle of radius 4 with component standard deviation 0.25. Bistable Langevin is the finite-time distribution generated by the double-well dynamics specified in Appendix B . Both targets are standardized before flow matching training.
Figure 6: Experiment 2, seed 1. As in the main-text representative, changing κ leaves the original learned-coordinate error unchanged (left), moves only the transverse amplification boundary (middle), and can leave the full augmented terminal error below tolerance even before transverse non-expansion is reached (right).
Figure 7: Experiment 2, seed 2. The same separation is reproduced in the third independently trained checkpoint for each target. These two appendix figures, together with Fig. 3 , underlie the three-seed summary in Tables 2 and 4 .
Coarse steps
Fine steps
Relative endpoint disagreement
16
32
3.3481×10−2
32
64
1.7390×10−2
64
128
5.4245×10−3
128
256
3.2001×10−3
256
512
9.6617×10−4
512
1024
3.7975×10−4
Appendix
Table 5: Task-00 RK4 endpoint refinement. Values are normalized by the finer endpoint norm.
N
1
2
3
4
5
6
7
8
9
10
RN
.0098
.4964
.6641
.7480
.7984
.8322
.8559
.8739
.8879
.8992
Appendix
Table 6: Task-00 candidate maxima RN on the reference-state Euler grids.
Task
Input
Gmax
tmax
χ
Residual
r⊥
r7
00
00
0.511522
0.03125
16.36869
0.00520
0.99978
0.88609
00
01
0.513789
0.03125
16.44125
0.00446
0.99997
0.90848
00
02
0.511219
0.03125
16.35901
0.00549
0.99968
0.28063
01
00
0.525207
0.03125
16.80663
0.00463
0.99972
0.96933
01
01
0.511849
0.03125
16.37918
0.00612
0.99991
0.43778
01
02
0.519417
0.03125
16.62134
0.00458
0.99996
0.97099
Appendix
Table 7: Maximum sampled G=tχ← for all 30 policy inputs. r7 is the norm fraction in the first seven internal coordinates of each action token.
Figure 8: Expanded view of the maximum sampled stability measure for the 30 policy inputs. The sufficient boundary G=1 lies outside this restricted vertical range and is shown in Figure 4 .
Input
t
Arnoldi χ
Full Jacobian χ
Relative difference
Task 00 / 00
1
0.50535361
0.50603892
0.1354%
Task 00 / 00
1/2
1.01024206
1.01169466
0.1436%
Task 00 / 00
1/32
16.36869355
16.40657338
0.2309%
Task 06 / 00
1/32
16.54621714
16.55066042
0.0268%
Appendix
Table 8: Arnoldi versus full-Jacobian local stability scales.
Symbol
Meaning
c
Fixed policy input: visual observation, robot state (proprioception), and language instruction.
ξ
Fixed Gaussian 10×32 source-noise tensor that initializes the action-flow ODE; treated separately from the policy input c .
t
Native OpenPI sampling time: 1 (noise) to 0 (action).
Jv
State Jacobian of the native-time velocity at fixed c,t .
EN(c,ξ)
Terminal error for one conditioning input and source-noise draw relative to the refined endpoint.
Nacc
Persistent trajectory-accuracy threshold within a finite tested budget set.
Lawrence Berkeley National Laboratory · International Computer Science Institute · 1Lawrence Berkeley National Laboratory, 2International Computer Science Institute +4