Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware

Yeonseok Lee

cs.LONov 3, 2026

Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware

Authors: Yeonseok Lee

Organizations: SLING AI Inc., Incheon, Republic of Korea

Abstract

Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders need sequential forward passes in proportion to the action horizon, each bounded below on embedded aerial platforms by streaming billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for 3D navigation on edge hardware. A compact RGB-D Vision Transformer and a five-head ensemble predictor, trained without pixel reconstruction, forecast the next latent state; a Conditional Flow Matching solver generates a velocity and yaw-rate chunk in two Euler steps, so each decision takes five sequential network evaluations; and the ensemble dispersion drives an uncertainty fallback. Version 1 specified the architecture, deployment and protocol without results. This version reports the first closed-loop evaluation: PX4 software-in-the-loop simulation of quadrotor missions in procedurally generated forests. Behaviour cloning failed on all 60 missions despite open-loop errors of 0.02-0.04 m/s: the policy copies its last command, a copycat form of causal confusion. DAgger with a privileged, safety-gated supervisor gave the first successes but plateaued at 11.1-21.1% across six checkpoints. A map-free runtime shield with a 15 s depth memory, barrier constraints, a cap on motion outside the 73-degree field of view and backoff along the flown path raised the first DAgger policy's success on 180 paired missions from 21.1% to 52.2% (95% CI [45.0, 59.4]; exact McNemar p = 1.2 x 10^-10) and cut its crashes from 72.2% to 22.2%, while timeouts rose from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points, at mean speeds of 1.16-1.36 m/s. The code (https://github.com/ylee-sling/Aero-VPA), the trained policy (72.6 MB, https://huggingface.co/ylee-sling/aero-vpa) and the full benchmark dataset are public.

Figures & tables

Panel of Figure 1 — Figure 1: Aero-VPA deployment architecture. The inference loop (top) runs once per camera frame and produces a velocity and yaw-rate chunk $\mathcal{C}_{t}$ through five sequential network evaluations; the safety filter applies the uncertainty-dependent attenuation $\alpha(\bar{\sigma})$ , the braking envelope $v_{\mathrm{safe}}$ , and the validity horizon $H_{t}$ . The executor (bottom) evaluates the newest valid chunk at its chunk-local time, converts it to NED setpoints, and streams them to PX4 over micro-XRCE-DDS; it engages the braking fallback when a fallback is requested or no valid chunk is available. The runtime shield (orange, added in version 2) filters every setpoint against a 15 s depth memory or flies recovery manoeuvres itself; the setpoint it dispatches is fed back as $\bar{a}_{t}$ (dashed). In the evaluation of Section 6 the companion-computer software ran on a cloud GPU against PX4 SITL (Section 4.6 ).

Panel of Figure 2 — Figure 2: Mode automaton of the runtime shield. In Fly the policy’s executor runs and the shield filters each setpoint. In Backoff , Nudge and Hold (orange) the executor is paused, and the shield’s own setpoint, passed through the same filters, is dispatched and fed back as $\bar{a}_{t}$ . The radius $r=0.55$ m is the barrier’s zero; a backoff that ends without an active waypoint returns to Fly .

Module	Configuration	Params	Evals./decision	FP16 weights
Encoder $E_{\psi}$	ViT, $4{\times}128{\times}160$ input, patch 16, width 384, depth 6, 6 heads	11.17M	1	22.3 MB
Selector $\pi^{h}_{\phi}$	MLP $785\!\to\!512\!\to\!512\!\to\!6$	0.67M	1	1.3 MB
Predictor $P_{\omega}$	5 heads, MLP $292\!\to\!512\!\to\!512\!\to\!256$	2.72M	1 (heads batched)	5.4 MB
Solver $v_{\theta}$	Transformer, width 256, depth 4, 4 heads, $H{+}1=17$ tokens	3.57M	$K=2$	7.1 MB
Trainable total		18.13M	$K{+}3=5$ sequential	36.3 MB
Text encoder (frozen)	CLIP ViT-B/32 text tower	63.4M	once per mission	–

Table 1: Aero-VPA network configuration and parameter budget (analytic counts for

H=16

d=256

N_{u}=6

; FP16 weights). The frozen text encoder is evaluated once per mission and lies outside the per-decision dependency chain.

			SR	CR	$\bar{v}$	$T_{\mathrm{e2e}}$ p50 / p95	$\bar{P}_{\mathrm{CGC}}$	$E_{\mathrm{dec}}$
ID	Method	Params	(%)	(%)	(m/s)	(ms)	(W)	(mJ)
B1	VINS-Fusion + OctoMap + min-snap	–	–	–	–	– / –	–	–
B2a	Vision PPO	matched	–	–	–	– / –	–	–
B2b	Vision SAC	matched	–	–	–	– / –	–	–
B3a	Diffusion Policy (DDIM 10 / 100)	matched	–	–	–	– / –	–	–
B3b	Octo-Small, fine-tuned	27M	–	–	–	– / –	–	–

Table 2: Target comparison schema (template; no results). Each block is reported per environment class (E1–E3), condition (ID/OOD), and speed limit

\bar{v}_{h}\in\{3,5,7\}

m/s. “–” denotes a quantity to be measured under the protocol of Section 5 ;

\bar{P}_{\mathrm{CGC}}

is the mean power of the VDD_CPU_GPU_CV rail. The last row states the acceptance criteria for Aero-VPA (V1) and the hypothesis each criterion tests.

ID	Factor	Levels	Primary readouts	Tests
A1	CFM integration steps	$K\in\{1,\underline{2},4,8\}$	$Kt_{v}$ , $T_{\mathrm{VPA}}$ , SR, CR, chunk error vs. $K=64$ reference	H 1 , H 2
A2	Uncertainty threshold $\times$ attenuation floor	$p_{\mathrm{fa}}\in\{10^{-2},10^{-3},\underline{2.8{\times}10^{-5}}\}$ or fallback off; $\times$ $\alpha_{\min}\in\{0.25,\underline{0.5},1\}$	Pareto, spurious and true fallback rates	H 3
A3	Attenuation law	none, piecewise linear (Eq. ( 24 )) , exponential	CR, $\bar{v}$ , jerk	H 3
A4	Braking envelope	off, on	CR in E1/E2, $\bar{v}$ , conservatism	H 3
A5	Noise persistence	fresh noise, persistent, $N_{\xi}=30$ , persistent, $N_{\xi}=10$	discontinuity at switches, SR in E3	H 5
A6	Latent predictor	removed (reactive CFM), JEPA + VICReg , JEPA without $v(Z),c(Z)$	SR, effective rank of $C(Z)$ , AUROC	H 4

Table 3: Planned ablation matrix. Defaults are underlined; all other factors are held at their defaults. “Pareto” denotes the front of collision rate versus mission time normalized by that of V0.

Policy	Training data	Inputs and recipe
BC NR	demos	non-residual stage (i); all inputs
BC	demos	residual stage (i); all inputs
BC $-\bar{a}$	demos	as BC, with $\bar{a}_{t}$ withheld
D 1	demos + DAgger round 1 (flown by BC)	all inputs
D 2	demos + rounds 1–2 (round 2 flown by D 1 )	all inputs
D 3 $-\bar{a}$	demos + rounds 1–3 (round 3 flown by D 2 )	$\bar{a}_{t}$ withheld

Table 4: Evaluated policies. Except BC NR , all use the frozen residual encoder

E_{\psi}

and predictor

P_{\omega}

; all are calibrated with the episode-maximum method of Section 3.5 . Withheld inputs enter

\pi^{h}_{\phi}

and

e_{t}

at their training mean;

P_{\omega}

always receives

\bar{a}_{t}

. “Demos”: demonstrations from the 200 training worlds.

Expert frames
Policy	$\bar{a}_{t}$ swap	$v^{\mathcal{H}}_{t}$ swap	image	goal	all of $S_{t}$	$\bar{a}_{t}$ to $P_{\omega}$ only	$\|v_{0}-v_{0}^{\star}\|$	selector
BC	0.85 (0.31)	0.31 (0.86)	0.01	0.01	1.14	0.00	0.02	98.0%
BC $-\bar{a}$	–	1.17 (0.21)	0.03	0.02	1.14	0.00	0.04	93.6%
D 1	0.96 (0.15)	0.17 (1.02)	0.02	0.01	–	–	–	–
D 2	0.93 (0.27)	0.31 (0.99)	0.02	0.01	–	–	–	–
D 3 $-\bar{a}$	–	1.10 (0.28)	–	–	–	–	–	–

Table 5: Copycat diagnostics. Top: 5000 held-out expert frames (validation flights). A swap replaces one input by that of another frame; entries are the median change of the first knot

v_{0}

(m/s), with the error to the donor frame’s label in parentheses (the median label difference between frames is 1.13 m/s). Open loop: median

|v_{0}-v_{0}^{\star}|

and selector agreement with the expert’s primitive. Bottom: the policies’ own DAgger states, labelled by the expert and held out for the policies shown; “

>45^{\circ}

”: share of frames whose command points more than 45° from the label. “–”: withheld input, or not measured.

Round	Flown by	Missions	Success	Crash	Policy’s share of the flight
1	BC	201	148 (73.6%)	28 (13.9%)	$\approx$ 50%
2	D 1	201	166 (82.6%)	5 (2.5%)	63%
3	D 2	201	189 (94.0%)	10 (5.0%)	64%

Table 6: DAgger rounds under the privileged supervisor (

N=201

missions per round over the training worlds). These flights are gated and partly flown by the expert, so their success and crash rates measure the supervised system, not autonomous performance.

Comparison (A vs B)	Only A succeeded	Only B succeeded	$p$
D 1 vs D 2	29	14	0.0315
D 1 vs D 3 $-\bar{a}$	28	10	0.0051
D 1 vs D 3 $-\bar{a},v$	33	19	0.0704
D 1 vs D 3,SD	28	18	0.184
D 1 vs D 3,KW	26	16	0.164
D 2 vs D 3 $-\bar{a}$	19	16	0.736

Table 7: Paired success comparisons without the shield (

N=180

paired test missions; last row

N=60

, the missions the BC policies flew). Discordant pairs and the two-sided exact McNemar

p

Checkpoint	U	S [95% CI]	$\Delta$ (pp)	both / U / S / neither	$p$	U $\to$ S	U $\to$ S
	Success			Paired success		Crash (%)	Timeout (%)
D 1	38 (21.1%)	94 (52.2%) [45.0, 59.4]	$+31.1$	26 / 12 / 68 / 74	$1.20\times 10^{-10}$	72.2 $\to$ 22.2	2.8 $\to$ 23.3
D 2	23 (12.8%)	77 (42.8%) [35.8, 50.1]	$+30.0$	14 / 9 / 63 / 94	$4.18\times 10^{-11}$	80.0 $\to$ 38.3	1.1 $\to$ 16.7
D 3 $-\bar{a}$	20 (11.1%)	79 (43.9%) [36.8, 51.2]	$+32.8$	11 / 9 / 68 / 92	$2.45\times 10^{-12}$	82.2 $\to$ 36.1	2.8 $\to$ 14.4
D 3 $-\bar{a},v$	24 (13.3%)	81 (45.0%) [37.9, 52.3]	$+31.7$	11 / 13 / 70 / 86	$1.33\times 10^{-10}$	77.2 $\to$ 42.8	3.3 $\to$ 10.0
D 3,SD	28 (15.6%)	82 (45.6%) [38.4, 52.8]	$+30.0$	19 / 9 / 63 / 89	$4.18\times 10^{-11}$	76.1 $\to$ 35.0	3.3 $\to$ 16.1

Table 8: Main result: the six DAgger checkpoints without and with the runtime shield on the same

N=180

paired test missions. Success counts with Wilson 95% intervals;

\Delta

: change in percentage points; paired: missions in which both, only the unshielded (U), only the shielded (S), or neither run succeeded, with the two-sided exact McNemar

p

. Rows follow the training lineage, not a ranking: no shielded checkpoint differs significantly from D 1 + shield (Table 17 ).

Panel of Figure 3 — Figure 3: Mission endings of the six DAgger checkpoints without and with the runtime shield ( $N=180$ paired test missions each); the number at the right of each bar is its success rate. The shield turns most collisions into successes or timeouts. Exact counts are in Table 15 . Six first-person-view (FPV) videos on the model card ( Code and Data Availability ) show D 1 + shield and D 3,KW + shield on one mission per density that both completed in the benchmark; they are re-flights for illustration, and the shares shown come from the benchmark runs.

	Dense		Medium		Sparse
Checkpoint	U	S [95% CI]	U	S [95% CI]	U	S [95% CI]
D 1	4 (6.7%)	28 (46.7%) [35, 59]	17 (28.3%)	29 (48.3%) [36, 61]	17 (28.3%)	37 (61.7%) [49, 73]
D 2	0 (0.0%)	21 (35.0%) [24, 48]	9 (15.0%)	32 (53.3%) [41, 65]	14 (23.3%)	24 (40.0%) [29, 53]
D 3 $-\bar{a}$	2 (3.3%)	23 (38.3%) [27, 51]	6 (10.0%)	25 (41.7%) [30, 54]	12 (20.0%)	31 (51.7%) [39, 64]
D 3 $-\bar{a},v$	1 (1.7%)	17 (28.3%) [19, 41]	6 (10.0%)	28 (46.7%) [35, 59]	17 (28.3%)	36 (60.0%) [47, 71]
D 3,SD	2 (3.3%)	17 (28.3%) [19, 41]	9 (15.0%)	26 (43.3%) [32, 56]	17 (28.3%)	39 (65.0%) [52, 76]

Table 9: Success by obstacle density, unshielded (U) and shielded (S), as successes of

N=60

missions per density (12 worlds

\times

5 missions), with Wilson 95% intervals (%) for the shielded runs. Trunk spacing: dense 3.5 m, medium 5 m, sparse 7 m.

Checkpoint	unshielded	shielded [95% CI]	both / only U / only S / neither	$p$	U $\to$ S	$p$
	Success		Paired success		Crashes
D 1	31	75 (52.1%) [44.0, 60.1]	21 / 10 / 54 / 59	$2.00\times 10^{-8}$	103 $\to$ 33	$1.52\times 10^{-15}$
D 2	20	64 (44.4%) [36.6, 52.6]	13 / 7 / 51 / 73	$2.40\times 10^{-9}$	112 $\to$ 53	$3.85\times 10^{-13}$
D 3 $-\bar{a}$	15	62 (43.1%) [35.3, 51.2]	8 / 7 / 54 / 75	$4.32\times 10^{-10}$	120 $\to$ 51	$9.43\times 10^{-16}$
D 3 $-\bar{a},v$	22	67 (46.5%) [38.6, 54.7]	10 / 12 / 57 / 65	$3.74\times 10^{-8}$	109 $\to$ 62	$3.81\times 10^{-8}$
D 3,SD	24	65 (45.1%) [37.2, 53.3]	15 / 9 / 50 / 70	$5.26\times 10^{-8}$	107 $\to$ 50	$1.26\times 10^{-12}$

Table 10: Missions never used to revise the shield: missions 1–4 of every test world (

N=144

paired missions per checkpoint; mission 0 of every world was the shield pilots’ set). Same worlds as the full benchmark, so this is a partial check, not a fresh-world test. Paired success counts and exact McNemar

p

; crashes unshielded

\to

shielded with the exact McNemar

p

of the paired crash comparison.

	D 1 , unshielded	+ shield v1	+ shield v2
Success [95% CI]	7 (19.4%) [10, 35]	12 (33.3%) [20, 50]	20 (55.6%) [40, 70]
Crashes (collision / ground)	27 (23 / 4)	11 (10 / 1)	9 (7 / 2)
Timeouts / camera timeouts	0 / 2	12 / 1	7 / 0
Paired with the unshielded run (both / only U / only S)	–	2 / 5 / 10, $p=0.30$	5 / 2 / 15, $p=0.0023$
Crashes while the shield flew	–	10 of 11	7 of 9
Collisions in view / beside / behind	13 / 12 / 0 of 25 †	2 / 4 / 4 of 10	6 / 1 / 0 of 7

Table 11: Shield design iteration on the pilot missions: mission 0 of every test world flown by D 1 (

N=36

). Paired: missions in which both, only the unshielded or only the shielded run succeeded, with the exact McNemar

p

. Crash locations give the bearing of the nearest trunk at the collision relative to the camera axis: in view within 36.5°, beside up to 100°, behind beyond. “Moving outside the view”: horizontal motion more than 36.5° off the camera axis, among collisions with a measurable direction (horizontal speed above 0.2 m/s). † From a separate traced flight of the same 36 missions without the shield (6 successes, 25 collisions, 2 ground contacts, 2 timeouts, 1 camera timeout).

	Per mission				Seconds per mission						Memory	Update
Checkpoint	backoffs	nudges	stalls	holds	barrier	blind	backoff	nudge	altitude	stale	clear. (m)	p95 (ms)
D 1	0.87	2.05	1.35	2.42	12.9	13.6	3.7	3.0	0.3	0.1	0.42	2.9
D 2	1.03	1.98	1.26	1.59	11.8	14.1	3.7	3.1	0.1	0.1	0.32	2.9
D 3 $-\bar{a}$	1.06	2.07	1.32	1.66	12.6	17.5	4.1	3.1	0.1	0.1	0.37	2.9
D 3 $-\bar{a},v$	0.91	1.72	1.11	1.35	11.1	14.4	2.6	2.6	0.0	0.0	0.33	3.0
D 3,SD	0.96	2.22	1.49	1.19	11.9	16.5	4.0	3.2	0.0	0.0	0.36	2.9

Table 12: Shield activity in the shielded benchmark (

N=180

missions per checkpoint): interventions per mission (means), seconds per mission with each rule active (means), clearance to the memory (median over missions of the per-mission minimum) and the memory update’s 95th percentile (median over missions). The uncertainty cap was disabled (0.0 s for every checkpoint).

			Medians				Per mission			Seconds
	Ending	$n$	waypoints	time (s)	speed (m/s)	clear. (m)	backoffs	nudges	stalls	barrier / blind
D 1	success	94	1.00	83.1	1.24	0.60	0.50	1.36	0.86	11.7 / 11.4
	timeout	42	0.71	120.3	0.85	0.53	1.17	4.26	3.10	17.7 / 25.0
	collision	33	0.25	50.3	1.13	0.35	1.52	1.27	0.67	11.4 / 7.7
	ground	7	0.50	56.0	1.08	0.40	1.43	2.57	1.14	12.8 / 8.1
	camera timeout	4	0.17	28.2	1.26	1.50	0.00	0.50	0.50	1.7 / 5.0

Table 13: What the shielded missions did, by ending, for D 1 and D 3,KW (

N=180

missions each;

n

: missions with that ending). Waypoint fraction, duration (time), mean speed and smallest privileged clearance are medians over the missions; interventions and the seconds with the barrier and the blind cap active are means per mission.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Parameter	Value
Observation	RGB-D frame; camera	$4\times 128\times 160$ ; HFOV 1.274 rad (73°), depth 0.1–10 m
	proprioception $S_{t}$ ; goal $g_{t}$	$\mathbb{R}^{13}$ ; $\mathbb{R}^{4}$ , distance saturated at 20 m
	waypoint acceptance $\tau_{w}$ ; instruction	1 m; CLIP ViT-B/32 text, 512-d (null embedding)
Action	chunk $H$ , knot spacing $\Delta$	16 knots, 0.02 s (0.32 s)
	bounds	7 m/s horizontal, 2 m/s vertical, 1.5 rad/s
$E_{\psi}$ (11.17M)	ViT	patch 16, 80 tokens + CLS, width 384, depth 6, 6 heads, $z\in\mathbb{R}^{256}$

Appendix

Table 14: Values used in every reported flight.

Checkpoint		Success	Collision	Ground	Timeout	Camera timeout	Out of bounds
D 1	U	38	114	16	5	7	0
	S	94	33	7	42	4	0
D 2	U	23	133	11	2	10	1
	S	77	64	5	30	4	0
D 3 $-\bar{a}$	U	20	131	17	5	7	0
	S	79	55	10	26	7	3

Appendix

Table 15: Exact ending counts of the six DAgger checkpoints, unshielded (U) and shielded (S), on the same

N=180

test missions.

Checkpoint	Crashes U	Crashes S	Only U crashed	Only S crashed	$p$
D 1	130	40	98	8	$8.08\times 10^{-21}$
D 2	144	69	83	8	$7.58\times 10^{-17}$
D 3 $-\bar{a}$	148	65	93	10	$5.20\times 10^{-18}$
D 3 $-\bar{a},v$	139	77	78	16	$5.78\times 10^{-11}$
D 3,SD	137	63	83	9	$3.92\times 10^{-16}$
D 3,KW	145	47	107	9	$2.00\times 10^{-22}$

Appendix

Table 16: Paired crash comparisons, shielded against unshielded (

N=180

paired missions): crashes in each condition, discordant pairs and the two-sided exact McNemar

p

Shielded checkpoint	counts	$p$	counts	$p$
	Against unshielded D 1		Against D 1 + shield
D 1	26 / 12 / 68 / 74	$1.20\times 10^{-10}$	–	–
D 2	17 / 21 / 60 / 82	$1.69\times 10^{-5}$	46 / 48 / 31 / 55	0.0712
D 3 $-\bar{a}$	17 / 21 / 62 / 80	$7.51\times 10^{-6}$	50 / 44 / 29 / 57	0.101
D 3 $-\bar{a},v$	19 / 19 / 62 / 80	$1.77\times 10^{-6}$	52 / 42 / 29 / 57	0.154
D 3,SD	22 / 16 / 60 / 82	$3.85\times 10^{-7}$	47 / 47 / 35 / 51	0.224

Appendix

Table 17: Each shielded checkpoint paired with unshielded D 1 and with D 1 + shield (

N=180

paired missions; both / only D 1 / only the other / neither, two-sided exact McNemar

p

Metric	D 1	D 2	D 3 $-\bar{a}$	D 3 $-\bar{a},v$	D 3,SD	D 3,KW
Waypoints reached	38.5%	28.9%	27.6%	31.3%	29.8%	31.9%
Duration / path (s, m)	38.7 / 46.1	28.9 / 35.1	30.6 / 33.3	32.7 / 43.5	30.1 / 41.4	35.5 / 38.6
Fallback onsets per mission	0.77	0.51	1.04	0.68	0.69	0.86
Missions with a fallback	21.1%	21.1%	30.6%	23.3%	26.1%	31.1%
Braking or holding (s per mission)	0.1	0.1	0.2	0.1	0.1	0.2
Time / expert reference	1.15	1.00	1.26	1.05	1.10	1.10

Appendix

Table 18: Secondary metrics without the shield (

N=180

missions per checkpoint). Duration, path, time and path ratios, smallest clearance and timing are medians over missions (time and path ratios over successful missions); waypoints and blind decisions are pooled shares; fallback figures and mean speed are means over missions.

Metric	D 1	D 2	D 3 $-\bar{a}$	D 3 $-\bar{a},v$	D 3,SD	D 3,KW
Waypoints reached	76.6%	68.4%	65.9%	62.2%	68.9%	72.9%
Duration / path (s, m)	86.3 / 95.4	77.9 / 88.0	76.1 / 90.5	60.7 / 81.2	77.6 / 90.0	82.2 / 93.5
Fallback onsets per mission	1.53	1.13	1.39	1.04	1.32	1.25
Missions with a fallback	30.0%	32.2%	34.4%	33.9%	36.7%	34.4%
Braking or holding (s per mission)	0.3	0.2	0.2	0.1	0.2	0.2
Time / expert reference	1.26	1.22	1.27	1.12	1.22	1.22

Appendix

Table 19: Secondary metrics with the shield (

N=180

missions per checkpoint; definitions as in Table 18 ).

Checkpoint	Dense	Medium	Sparse
D 1	28 / 17 / 15	29 / 16 / 15	37 / 7 / 12
D 2	21 / 29 / 10	32 / 19 / 9	24 / 21 / 11
D 3 $-\bar{a}$	23 / 24 / 12	25 / 28 / 6	31 / 13 / 8
D 3 $-\bar{a},v$	17 / 31 / 12	28 / 30 / 2	36 / 16 / 4
D 3,SD	17 / 31 / 12	26 / 21 / 11	39 / 11 / 6
D 3,KW	22 / 20 / 18	33 / 12 / 15	35 / 15 / 6

Appendix

Table 20: Shielded endings by density: successes / crashes / timeouts of

N=60

missions per density; the remainder of each 60 are camera timeouts or out-of-bounds endings.

Policy	Stage (i)	Success	Crash (collision / ground)	Other endings
BC NR	non-residual	0/60	83% (48% / 35%)	10 missions, not broken down
BC	residual	0/60	51.7% (42% / 10%)	timeout 48.3%
BC $-\bar{a}$	residual	0/60	70.0% (22% / 48%)	timeout 8.3%, camera timeout 21.7%

Appendix

Table 21: Behaviour-cloning policies in closed loop (

N=60

missions each: 12 of the 36 test worlds, 5 missions per world, flown before the warm-up fix). Crash shares are split into collision and ground contact.

	D 3,SD				D 3,KW
Swapped input	change	primitive	own	donor	change	primitive	own	donor
none	0.00 / 0.00	0.0%	0.04	1.14	0.00 / 0.00	0.0%	0.03	1.14
image $z_{t}$	0.04 / 0.13	12.2%	0.04	1.14	0.03 / 0.08	10.4%	0.03	1.15
goal $g_{t}$	0.03 / 0.40	16.9%	0.05	1.09	0.02 / 0.09	9.0%	0.03	1.14
velocity $v^{\mathcal{H}}_{t}$	0.16 / 1.54	12.9%	0.18	0.98	0.49 / 1.85	28.9%	0.50	0.97
attitude $n^{\mathcal{B}}_{t}$	0.05 / 0.89	17.3%	0.07	1.16	0.04 / 0.57	12.3%	0.05	1.16

Appendix

Table 22: Swap tests of D 3,SD and D 3,KW on

N=5000

held-out expert frames: change of the first knot

v_{0}

(median / p90, m/s), share of frames whose primitive changed, and the error to the frame’s own label and to the donor frame’s label (m/s). “

\bar{a}_{t}=0

”: the last command set to zero (first-knot speed afterwards in parentheses). Open loop on the same frames:

|v_{0}-v_{0}^{\star}|

median / p90 0.04 / 0.10 m/s (D 3,SD ) and 0.03 / 0.07 m/s (D 3,KW ); selector agreement 91.8% and 92.9%; direction error above

45^{\circ}

in 0.6% and 0.4% of frames.

UAV Navigation

Unmanned aerial vehicle (UAV) navigation develops methods for locating and guiding drones safely through known, unknown, and changing environments.

See all 277 papers

Apr 27, 2026

LiteVLA-H: Dual-Rate Vision-Language-Action Inference for Onboard Aerial Guidance and Semantic Perception

Vision-language-action (VLA) models have shown strong semantic grounding and task generalization in manipulation, but aerial deployment remains difficult because drones require low-latency closed-loop guidance under strict onboard compute and communication constraints. We present LiteVLA-H, a compact 256M-parameter VLA system designed for dual-rate operation on an NVIDIA Jetson AGX Orin: a fast outer-loop guidance mode for short action-token outputs and a slower semantic mode for scene understanding, hazard description, and operator-facing narration. The central empirical observation is that, in this compact edge regime, end-to-end latency is dominated by multimodal pre-fill rather than by the marginal cost of decoding a few extra tokens. This motivates a scheduler that issues reactive action tokens at 50.65,ms (19.74,Hz) while still supporting sentence-level semantic outputs at 149.90--164.57\ms (6.08--6.67,Hz) on the same embedded platform. To specialize the model without collapsing its descriptive competence, we use a knowledge-preserving fine-tuning recipe that mixes reactive flight data, aerial semantic data, and generic caption/VQA supervision. Beyond reporting current latency measurements, we position the system against recent state-of-the-art architectures, including AnywhereVLA, FutureVLA, and ReMem-VLA, showing that the measured action branch reaches a higher edge inference rate under our deployment conditions while retaining periodic semantic awareness.

Justin williams, Kishor Datta Gupta, Roy George +1

Aerial Robotics Efficient VLM Inference Vision-Language-Action Models

Ask PDF

Oct 11, 2026

Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors

Vision-Language-Action (VLA) models commonly generate robot actions token by token through Large Language Model (LLM) decoders, causing inference latency to grow with action sequence length and limiting closed-loop control frequency. We present Goal-Conditioned Vision-Predictor-Action (VPA), which replaces autoregressive decoders with a lightweight latent control loop. A shared vision encoder and a frozen text encoder embed observations, goal images, and instructions; a neuro-symbolic selector chooses a discrete skill; a Joint-Embedding Predictive Architecture (JEPA) predictor forecasts the next latent state; and a flow-matching solver generates an entire action chunk in a few integration steps, so that sequential network evaluations per decision step do not depend on the action horizon. We analyze inference complexity, regularization against representation collapse, and conditions under which latent sub-goal thresholding limits error accumulation across task phases. As an initial proof of concept, we evaluated VPA on the LIBERO-10 suite. The trainable model has under 20M parameters (checkpoint under 75 MB, excluding a frozen 63.4M-parameter CLIP text encoder); a decision step of a single-camera reference configuration takes 10-14 ms on an Apple M4; and trained only on LIBERO-10 demonstrations in under three hours on a single GPU, a two-camera policy succeeded in 40.0% of 200 evaluation episodes. These preliminary results from a single training run suggest the approach is workable. Code and weights are available at https://github.com/ylee-sling/Vision-Predictor-Action and https://huggingface.co/ylee-sling/vpa-libero10, and archived at https://doi.org/10.5281/zenodo.22960920.

Yeonseok Lee

SLING AI Inc., Incheon, Republic of Korea

Flow Matching for Action Generation Efficient VLA Models Vision-Language-Action Models Robotic Manipulation Joint-Embedding Predictive Architecture

Ask PDF

Oct 18, 2026

AeroDASH: Asynchronous Vision-Language-Action Control for Continuous UAV Flight

Onboard inference latency creates a mismatch between the trajectories predicted by vision-language-action (VLA) policies and the ongoing motion of unmanned aerial vehicles (UAVs). We present AeroDASH, an asynchronous VLA framework for continuous language-conditioned flight. A B-spline policy is initialized through supervised fine-tuning on expert demonstrations, providing continuous trajectory references in position and yaw. During execution, trajectory fusion removes the expired prediction prefix and connects the remaining segment to the active reference with derivative continuity while limiting trajectory deformation. To account for deviations induced by asynchronous execution, execution-aware group relative policy optimization (GRPO) further adapts the policy using trajectory-level rewards. On the UAV-Flow test set, AeroDASH achieves a success rate of 76.56%, exceeding OpenVLA-UAV and AeroDASH w/o GRPO by 38.1 and 20.1 percentage points, respectively. Execution experiments show that AeroDASH sustains motion during inference and reduces acceleration, jerk, and speed fluctuations relative to asynchronous tracking of waypoint sequences. We further demonstrate language-conditioned flight with onboard inference and asynchronous trajectory execution. The code will be released upon acceptance.

Zhe Liu, Yuxuan Zheng, Jinzhe Qin +2

Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China · Department of Mechanical and Energy Engineering, Southern University of Science and Technology, Shenzhen, China · Differential Robotics

UAV Navigation RL for VLMs

Ask PDF

Abstract

UAV Navigation

Explore similar work

LiteVLA-H: Dual-Rate Vision-Language-Action Inference for Onboard Aerial Guidance and Semantic Perception

Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors

AeroDASH: Asynchronous Vision-Language-Action Control for Continuous UAV Flight