Authors: Mingyang Lyu, Yinqian Sun, Gongce Cao, Guobin Shen, Yuhao Zhang, Lu Jia, Chenfei Yan, Qian Zhang, Pin Lyu, Chen Gao, Bo Wang
Organizations: Institute of Automation, Chinese Academy of Sciences, Beijing, China · Zhongguancun Academy, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · ByteDance, China · University of Illinois Urbana-Champaign, USA · University of Science and Technology Beijing, Beijing, China · Beihang University, Beijing, China · Tsinghua University, Beijing, China
Vision-language-action (VLA) policies implicitly assume that third-person cameras remain fixed at their data-collection poses; once a camera is bumped, remounted, or swapped at deployment, task success degrades precipitously even though the underlying task and robot remain unchanged. We present Homeward, a training-free, plug-and-play observation adapter that canonicalizes observations from an uncalibrated deployment camera back into the training view for frozen VLA policies. Our design stems from a closed-loop diagnostic paradox in simulation: while re-rendering perturbed views into the training camera with oracle geometry recovers much of the lost performance, substituting an off-the-shelf feed-forward 3D foundation model can leave the policy worse off than applying no correction at all. This degradation concentrates on the thin, articulated robot arm, which feed-forward depth prediction tears and scatters - yet the arm's exact 3D state is already known from robot kinematics. Homeward exploits this asymmetry: it estimates what is unknown - camera parameters and static scene depth - from the live RGB frame and a single training-view reference, and renders what is known - the moving robot links - directly from the kinematic model. Evaluated across ∼300,000 closed-loop episodes spanning four frozen VLAs, Homeward boosts success across all 1,599 camera variants of LIBERO-Plus from 62.5%, 76.5%, 68.0%, and 3.6% to 77.6%, 95.1%, 91.2%, and 26.4% for OpenVLA-OFT, π0.5, π0, and OpenVLA, respectively. Without any scene fine-tuning or deployment camera calibration, Homeward extends to 6-DoF random placements, focal-length changes, and, with per-step pose estimation, continuous intra-episode camera drift, while matching or surpassing a calibrated, scene-trained view-synthesis baseline on three of four policies on LIBERO-Plus.
Figures & tables
Figure 2: Overview of Homeward . (a) Monocular scene restoration: At the first policy query, frozen VGGT estimates and caches the deployment camera transform ( C′→C⋆ ) from [Iref,I0′] ; subsequent queries infer single-frame depth Dt aligned on static pixels. Depth-edge filtering, z-buffer splatting, and disocclusion filling from Iref yield the scene layer St , which retains artifacts on the moving arm (inset). (b) Kinematic robot layer & policy execution: Forward kinematics and the one-time training camera calibration C⋆ render moving-link RGB Rt and silhouette mask Mt , composited via I^t=Mt⊙Rt+(1−Mt)⊙St and passed with the unchanged Wt , qt , and ℓ to the frozen VLA π .
trained or
moved
robot
fine-tuned
camera
region
AnyCamVLA [ 14 ]
synthesis model
calibrated
synthesized
VistaBot [ 12 ]
VGGT, policy
intrinsics
reconstructed
Mirage [ 4 ]
dynamics model
calibrated
rendered
Homeward
none
uncalibrated
rendered
Table 1: Deployment assumptions of test-time viewpoint adapters. Trained or fine-tuned lists modules trained beyond public weights; Moved camera indicates calibration required at deployment. All but VistaBot assume one-time calibration of the training camera C⋆ (Mirage targets cross-embodiment transfer).
OFT
π0.5
π0
OpenVLA
G
L
G
L
G
L
G
L
frozen policy
79.7
35.3
84.3
52.7
65.2
44.9
8.1
1.2
oracle re-rendering †
90.4 ∗
76.1 ∗
88.7
77.8 ∗
89.7 ∗
74.7 ∗
27.5 ∗
15.0 ∗
oracle homography †
70.8 ∗
38.7
74.8 ∗
51.3
60.8
41.5
13.7 ∗
4.5 ∗
scene restoration only
72.8 ∗
51.3 ∗
76.0 ∗
60.4 ∗
82.6 ∗
62.5 ∗
22.3 ∗
7.2 ∗
Homeward
91.4 ∗
72.3 ∗
94.4 ∗
94.0 ∗
93.4 ∗
78.3 ∗
39.2 ∗
11.2 ∗
Table 2: Closed-loop diagnostic study (success rate, %) on LIBERO-Plus camera variants of Goal (G, 408) and Long (L, 419). † Privileged, non-deployable (simulator depth and/or cameras). ∗ Exact McNemar p<0.05 vs . frozen policy.
OpenVLA-OFT
π0.5
π0
OpenVLA
Suite
Variants
frozen
+ Homeward
frozen
+ Homeward
frozen
+ Homeward
frozen
+ Homeward
Spatial
376
91.8
97.3 ∗
76.9
98.1 ∗
79.5
98.7 ∗
3.2
32.2 ∗
Object
396
45.7
50.3
93.2
93.9
84.6
95.7 ∗
2.0
23.7 ∗
Goal
408
79.7
91.4 ∗
84.3
94.4 ∗
65.2
93.4 ∗
8.1
39.2 ∗
Long
419
35.3
72.3 ∗
52.7
94.0 ∗
44.9
78.3 ∗
1.2
11.2 ∗
All
1,599
62.5
77.6 ∗
76.5
95.1 ∗
68.0
91.2 ∗
3.6
26.4 ∗
Table 3: LIBERO-Plus camera dimension. Success rate (%) of frozen policies on all 1,599 camera variants, one episode per variant, without and with Homeward . ∗ exact McNemar p<0.05 on the paired episodes; for “All”, p<10−31 for every policy. No policy is retrained and no calibration of the moved camera is given.
Figure 3: Closed-loop success across camera-shift severity levels , pooled across all four LIBERO-Plus suites (levels 1–5 comprise 199, 360, 325, 228, and 487 variants). While frozen policies degrade steeply with viewpoint shift, Homeward degrades far less up to Level 5—lifting OpenVLA-OFT from 22.0% to 53.8%, π0.5 from 55.6% to 92.6%, and π0 from 48.3% to 80.5%. OpenVLA benefits most at Level 1 (10.6% to 47.2%).
OFT
π0.5
π0
OpenVLA
CamRand
frozen
83.8
93.9
88.2
30.5
+ Homeward
93.6 ∗
97.2 ∗
93.0 ∗
62.8 ∗
AnyCamVLA-style, other suites
frozen
62.0
59.3
43.3
5.9
+ Homeward
70.9 ∗
79.7 ∗
62.3 ∗
12.6 ∗
Table 4: Other camera changes. Success rate (%) pooled over suites: CamRand on all four (1,200 episodes), AnyCamVLA on Goal, Spatial and Object (900), drift and lens change on all four (1,200 and 800). ∗ McNemar p<0.05 .
scene
moved
training
camera
OFT
π0.5
π0
OpenVLA
frozen policy
–
–
62.5
76.5
68.0
3.6
AnyCamVLA ‡
fine-tuned
true
80.7
85.1
87.5
26.0
AnyCamVLA §
fine-tuned
estimated
79.5
82.1
83.1
14.1
Homeward
none
estimated
77.6
95.1
91.2
26.4
Table 5: Comparison with calibrated view synthesis on all 1,599 LIBERO-Plus camera variants (success rate, %). AnyCamVLA runs with the true ( ‡ ) or VGGT-estimated ( § ) parameters of the moved camera. Best per column in bold; per-suite results are in the supplementary material.
OFT
OFT+ (fine-tuned)
frozen
+ ours
frozen
+ ours
LIBERO-Plus camera
62.5
77.6 ∗
95.7
91.1 ∗
AnyCamVLA, Long
33.7
61.7 ∗
68.2
83.3 ∗
AnyCamVLA-style, other suites
62.0
70.9 ∗
84.2
87.4 ∗
CamRand
83.8
93.6 ∗
91.2
95.4 ∗
Drift (per-step pose)
56.0
74.7 ∗
76.5
86.7 ∗
Table 6: Restoration on top of a fine-tuned policy. Success rate (%) pooled as in Tabs. 3 and 4 , except Drift (Goal, Long) and Lens change (Goal, Spatial), the suites released for OpenVLA-OFT+; drift uses the per-step variant, the clean camera the gated default. ∗ McNemar p<0.05 against the same policy frozen.
Figure 4: Qualitative canonicalization at matched simulator states for a large AnyCamVLA orbit (Long) and an FoV narrowing ( 45∘→35∘ ). In frames from episodes where the frozen policy fails and Homeward succeeds, view synthesis blurs or omits fine objects, whereas Homeward preserves sharp scene geometry and crisp manipulator links.
OFT
π0.5
π0
OpenVLA
Goal
Long
Goal
Long
Goal
Long
Goal
Long
frozen policy
79.7
35.3
84.3
52.7
65.2
44.9
8.1
1.2
scene restoration only
72.8 ∗
51.3 ∗
76.0 ∗
60.4 ∗
82.6 ∗
62.5 ∗
22.3 ∗
7.2 ∗
+ whole robot incl. pedestal
39.0 ∗
31.0
72.3 ∗
53.9
74.3 ∗
43.4
4.4 ∗
0.0
+ moving links, pose per step
88.2 ∗
70.6 ∗
93.6 ∗
83.1 ∗
90.7 ∗
74.0 ∗
36.3 ∗
12.2 ∗
+ moving links, pose once
90.0 ∗
71.4 ∗
93.4 ∗
92.6 ∗
92.9 ∗
73.5 ∗
35.8 ∗
11.0 ∗
Table 7: Ablation study. Success rate (%) on the LIBERO-Plus camera variants of Goal (408) and Long (419). Rows marked (per-step) change one component of the per-step variant. – : not run. ∗ McNemar p<0.05 against the frozen policy.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S2 : Episodes seen from the training camera , frozen OpenVLA-OFT (top row of each pair) and with Homeward (bottom row), at evenly spaced queries of each episode; one query is eight control steps. Top: LIBERO-Plus Goal, level 5 (the scene of the main paper’s first figure). Middle: LIBERO-Plus Long, level 5. Bottom: AnyCamVLA large orbit on Long.
LIBERO-Long
small
medium
large
all
OpenVLA-OFT
frozen
80.6
9.6
11.0
33.7
+ Homeward
88.4
62.6
34.0
61.7 ∗
+ AnyCamVLA ‡
89.6
79.4
61.6
76.9 ∗
+ AnyCamVLA §
85.6
78.8
57.8
74.1 ∗
+ zero-shot LVSM ‡
40.8
18.2
2.6
20.5 ∗
Appendix
Table S2: AnyCamVLA protocol on LIBERO-Long, 500 episodes per offset (10 tasks × 50 initial states). Our frozen OpenVLA-OFT and π0.5 reach 33.7% and 50.3%, within 1.3 points of the 32.9% and 49.0% reported in [ 14 ] . ‡ AnyCamVLA’s released view-synthesis model run in our harness with the true parameters of the moved third-person camera, the wrist camera and the training camera; it is fine-tuned on LIBERO-Plus renders. § The same model with the moved camera’s pose and intrinsics estimated by VGGT against a training-view reference frame, as Homeward estimates them; the metric scale comes from the training view’s depth, and the wrist and training cameras keep their true parameters. Zero-shot LVSM: the same pipeline with LVSM’s public RealEstate10K weights [ 17 ] , without AnyCamVLA’s fine-tuning. Homeward requires no supplied calibration of the moved or wrist camera. ∗ McNemar p<0.05 against the frozen policy.
All suites
Object
Policy
Homeward
+ refined erase
Homeward
+ refined erase
LIBERO-Plus camera variants
π0.5
95.1
95.4
93.9
97.0 ∗
π0
91.2
90.3
95.7
97.0
OpenVLA
31.6 †
27.7 ∗
23.7
12.1 ∗
AnyCamVLA protocol, all orbits
Appendix
Table S3: Erasing the real arm with a refined camera. Success rate (%) of Homeward without and with erasure of the real arm under a camera refined from the arm’s motion, on the same episodes. All suites: the 1,599 LIBERO-Plus camera variants or the 6,000 AnyCamVLA episodes; † Goal and Object only (804). ∗ McNemar p<0.05 against Homeward .
OFT
π0.5
π0
OpenVLA
Category
n
frozen
+ Homeward
frozen
+ Homeward
frozen
+ Homeward
frozen
+ Homeward
Background Textures
400
62.2
80.5 ∗
83.2
92.8 ∗
69.0
84.8 ∗
4.8
31.2 ∗
Language Instructions
400
58.2
78.0 ∗
75.2
88.5 ∗
53.0
65.2 ∗
3.0
23.8 ∗
Light Conditions
400
69.8
89.0 ∗
88.2
96.2 ∗
75.5
89.8 ∗
3.2
24.8 ∗
Objects Layout
400
46.5
67.2 ∗
76.5
84.5 ∗
58.2
75.5 ∗
4.5
34.2 ∗
Robot Initial States
400
23.2
34.2 ∗
66.5
80.5 ∗
36.2
48.2 ∗
2.5
11.8 ∗
Appendix
Table S4: Compound perturbations. Success rate (%) on the 2,400 non-camera variants of LIBERO-Plus (600 per suite), each under a CamRand camera at the largest radius, by perturbation category; one episode per variant. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0
Goal
Long
Goal
Long
Goal
Long
frozen policy
81.6
34.8
83.8
52.7
65.2
44.9
Homeward , exact training camera
87.3 ∗
71.1 ∗
95.6 ∗
92.4 ∗
92.9 ∗
73.5 ∗
camera off by 5 mm, 0.5∘
89.0 ∗
69.5 ∗
94.9 ∗
89.7 ∗
92.9 ∗
74.2 ∗
camera off by 10 mm, 1∘
88.7 ∗
68.0 ∗
93.4 ∗
90.9 ∗
89.7 ∗
75.9 ∗
camera off by 20 mm, 2∘
86.0
58.9 ∗
93.4 ∗
86.6 ∗
90.7 ∗
75.2 ∗
Appendix
Table S5 : Calibration error of the training camera. Success rate (%) on the LIBERO-Plus camera variants of Goal (408) and Long (419) when the training camera that the robot is drawn from is off by the given translation and rotation (random direction and axis, fixed per episode). All rows run on one machine, whose renderer differs slightly from that of the ablation table of the main paper, and without the depth-edge filter. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0
frozen policy
62.5
76.5
68.0
+ temporal ensembling of action chunks
64.7 ∗
74.8
67.3
+ 2D registration to the reference frame
62.4
74.1 ∗
70.7 ∗
+ Homeward
77.6 ∗
95.1 ∗
91.2 ∗
Appendix
Table S6: Generic test-time fixes. Success rate (%) on all 1,599 LIBERO-Plus camera variants, pooled over the four suites. Temporal ensembling queries the policy every 4 control steps and averages the overlapping action chunks [ 48 ] . 2D registration warps every frame onto the reference frame with a homography fitted once per episode to SIFT matches by RANSAC [ 27 , 9 ] ; like Homeward , it requires no supplied camera calibration. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0
Goal
Long
Goal
Long
Goal
Long
frozen policy
79.7
35.3
84.3
52.7
65.2
44.9
2D registration
78.4
44.6 ∗
84.8
56.1
70.8 ∗
49.6 ∗
+ robot layer
71.6 ∗
51.3 ∗
89.7 ∗
67.5 ∗
87.5 ∗
61.8 ∗
oracle homography †
70.8 ∗
38.7
74.8 ∗
51.3
60.8
41.5
+ robot layer
79.7
46.3 ∗
90.9 ∗
78.5 ∗
86.0 ∗
63.5 ∗
Appendix
Table S7: 2D alignment with the kinematic robot layer. Success rate (%) on the LIBERO-Plus camera variants of Goal (408) and Long (419). † uses the true cameras; not deployable. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0 †
OpenVLA
Goal
Long
Goal
Long
Goal
Long
Spatial
Object
frozen policy
79.7
35.3
84.3
52.7
64.5
45.6
3.2
2.0
exact render
91.4 ∗
72.3 ∗
94.4 ∗
94.0 ∗
91.7 ∗
73.3 ∗
32.2 ∗
23.7 ∗
S=25
91.9 ∗
67.1 ∗
92.6 ∗
91.4 ∗
94.6 ∗
78.5 ∗
31.6 ∗
20.7 ∗
S=50
89.2 ∗
64.2 ∗
92.2 ∗
90.2 ∗
91.9 ∗
77.1 ∗
28.7 ∗
15.4 ∗
S=100
87.5 ∗
57.3 ∗
90.7 ∗
88.3 ∗
90.7 ∗
81.4 ∗
28.7 ∗
13.4 ∗
Appendix
Table S8: Colour error of the robot layer. Success rate (%) on the LIBERO-Plus camera variants. † noise seed 9, paired with the frozen policy under the same seed. OpenVLA ran this study on Spatial and Object first. – : not run. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0
OpenVLA
Goal
Long
Goal
Long
Goal
Long
Goal
Long
Homeward , exact render
91.4 ∗
72.3 ∗
94.4 ∗
94.0 ∗
93.4 ∗
78.3 ∗
39.2 ∗
11.2 ∗
robot drawn with a real-robot colour error
89.2 ∗
64.2 ∗
92.2 ∗
90.2 ∗
91.7 ∗
79.5 ∗
30.4 ∗
10.0 ∗
+ colour fit on the reference frame
87.3 ∗
64.9 ∗
92.6 ∗
86.6 ∗
93.9 ∗
79.7 ∗
21.8 ∗
13.6 ∗
Appendix
Table S9: Colour error on the suites of the main ablation table. Success rate (%) on the LIBERO-Plus camera variants of Goal (408) and Long (419), in the layout of the ablation table of the main paper. ∗ McNemar p<0.05 against the frozen policy.
OFT
π0.5
π0
OpenVLA
clean camera, 400 episodes
frozen
94.8
96.8
92.5
75.2
+ Homeward
94.8
96.2
92.8
74.8
without motion gate
95.8
97.0
92.8
73.5
non-camera variants of the four suites, 2,400 episodes
frozen
74.5
90.7
76.0
34.1
Appendix
Table S10: Performance without a camera shift. Success rate (%) on 400 clean-camera episodes (100 per suite) and on the 2,400 LIBERO-Plus non-camera variants of the four suites (lighting, background, noise, layout, robot initial state, language; 600 per suite). ∗ McNemar p<0.05 .
small drift (400)
all drift (1,200)
frozen
no gate
first query
every query
frozen
no gate
first query
every query
OpenVLA-OFT
85.2
85.2
84.8
86.8
58.1
72.5
72.3
73.0
π0.5
84.8
90.0
84.8
88.0
67.5
82.5
80.8
80.8
π0
64.8
66.8
64.2
69.0
50.3
64.0
63.2
63.8
OpenVLA
30.8
23.8
30.2
–
11.4
12.4
14.6
–
Appendix
Table S11: Motion gate under drift. Success rate (%) of the per-step variant. First query: computed from the ungated and the frozen runs of the same episodes. Every query: run with the gate decided at each query; OpenVLA was not run. On medium and large drifts the first-query gate never fires.
OFT
π0.5
π0
drift (1,200)
Goal
Long
Goal
Long
Goal
Long
OFT
π0.5
π0
frozen policy
79.7
35.3
84.3
52.7
65.2
44.9
58.1
67.5
50.3
acting on 0.25 s-old observations
31.9 ∗
10.7 ∗
55.1 ∗
34.6 ∗
50.2 ∗
13.4 ∗
–
–
–
Homeward
91.4
72.3
94.4
94.0
93.4
78.3
72.5
82.5
64.0
robot holds 0.25 s per query
92.4
73.5
90.7 ∗
92.1
92.9
77.3
69.2 ∗
81.2
64.3
acting on 0.25 s-old observations
21.3 ∗
24.3 ∗
56.9 ∗
37.9 ∗
47.5 ∗
26.5 ∗
–
–
–
Appendix
Table S12: A 0.25 s delay per policy query. Success rate (%) on the LIBERO-Plus camera variants of Goal (408) and Long (419) and on the drift benchmark (per-step variant). – : not run. ∗ McNemar p<0.05 against the same arm without delay.
variants
rotation (deg)
position (cm)
focal (%)
position > 20 cm
Spatial
376
1.2 / 3.0
5.2 / 9.4
5.2 / 14.1
0
Object
396
0.9 / 1.7
5.8 / 64.4
3.5 / 7.9
53
Goal
408
1.2 / 2.3
5.9 / 8.8
6.6 / 12.8
0
Long
419
0.9 / 1.7
4.5 / 7.5
4.7 / 13.6
0
All
1,599
1.0 / 2.3
5.4 / 9.1
4.5 / 12.5
53
Appendix
Table S13: First-query camera estimate against the simulator’s camera on the LIBERO-Plus camera variants: median / 90th percentile of the rotation error, the position error and the relative focal-length error, and the number of variants with a position error above 20 cm.
position error ≤ 20 cm (343)
position error > 20 cm (53)
frozen
Homeward
oracle geometry + robot layer †
frozen
Homeward
oracle geometry + robot layer †
OpenVLA-OFT
51.9
57.1
98.8 ∗
5.7
5.7
69.8 ∗
π0.5
92.1
98.0 ∗
98.8 ∗
100.0
67.9 ∗
100.0
π0
91.5
96.5 ∗
96.5 ∗
39.6
90.6 ∗
96.2 ∗
Appendix
Table S14: Object suite by camera-estimate error. Success rate (%) on the LIBERO-Plus camera variants of Object. † simulator depth and cameras; not deployable. ∗ McNemar p<0.05 against the frozen policy.
Figure S3 : Failure cases of Homeward with OpenVLA-OFT. Columns: the moved camera’s frame, Homeward ’s output, the oracle re-rendering from the simulator’s depth and cameras (no robot layer), and the training camera’s view of the same simulator state. Rows 1 and 3 show the first policy query, rows 2 and 4 the fifth and ninth. Rotations are those estimated at the first query.
passed (%)
OFT
π0.5
π0
OpenVLA
gate thresholds
L1
all
L1
all
L1
all
L1
all
L1
all
no gate ( Homeward as reported)
0
0
87.4
77.6
98.0
95.1
97.0
91.2
47.2
26.4
0.5∘ , 1% (default)
0
0
87.4
77.6
98.0
95.1
97.0
91.2
47.2
26.4
2∘ , 4% (4 × )
0
0
87.4
77.6
98.0
95.1
97.0
91.2
47.2
26.4
5∘ , 10% (10 × )
3
2
86.9
77.4
98.0
95.4
96.5
90.6 ∗
46.7
26.3
10∘ , 20% (20 × )
31
14
85.9
76.0 ∗
98.0
95.4
96.0
88.6 ∗
27.6 ∗
19.8 ∗
Appendix
Table S15: Wider motion gates, emulated on the LIBERO-Plus camera variants. Success rate (%) at Level 1 (L1, 199 variants) and over all 1,599 variants when episodes whose first-query estimate is below both thresholds (rotation; translation relative to the median scene depth) are passed through unrestored. Computed from the frozen and the Homeward runs of the same episodes; no threshold was tuned. Passed: share of episodes below the thresholds in the OpenVLA-OFT runs (the other runs differ by at most 0.5 points). ∗ McNemar p<0.05 against no gate.
OFT
π0.5
π0
OpenVLA
Goal
Long
Goal
Long
Goal
Long
Goal
Long
training camera, frozen policy †
98.0
91.0
96.0
91.0
95.0
84.0
74.0
56.0
moved camera, frozen policy
79.7
35.3
84.3
52.7
65.2
44.9
8.1
1.2
oracle re-rendering ‡
90.4
76.1
88.7
77.8
89.7
74.7
27.5
15.0
oracle re-rendering + robot layer ‡
91.4
83.3
97.1
91.9
95.8
80.2
52.7
25.1
Homeward
91.4
72.3
94.4
94.0
93.4
78.3
39.2
11.2
Appendix
Table S16: Training camera and oracle re-rendering. Success rate (%). † Clean-camera episodes: the unperturbed scene seen by the training camera, 100 per suite (10 tasks × 10 initial states). Other rows: the LIBERO-Plus camera variants of Goal (408) and Long (419), as in the ablation table of the main paper. ‡ Privileged: simulator depth and cameras.
models
time per query (ms)
peak GPU memory (GiB)
Homeward
VGGT-1B
164 + 10
6.6
Homeward , per-step
VGGT-1B
191 + 10
6.6
AnyCamVLA, true cameras ‡
LVSM
32
1.1
AnyCamVLA, estimated cameras §
LVSM, VGGT-1B
31
7.2
Appendix
Table S17: Inference cost of view restoration, excluding the policy. Homeward : scene layer + robot layer. § VGGT runs once per episode, on the first frame; its time is not included.
frozen
Homeward
per-step
π0
GPU
software
GPU
software
GPU
software
Spatial
79.5
79.5
97.6
97.3
97.9
97.1
Object
84.6
85.1
96.0
96.2
96.7
96.2
Goal
65.2
64.7
92.9
93.1
90.7
91.9
Long
44.9
48.0
73.5
78.0 ∗
74.0
76.1
Appendix
Table S18: Software rendering. Success rate (%) of π0 on the LIBERO-Plus camera variants with MuJoCo rendered on the GPU (NVIDIA EGL, used for every other result) and in software (Mesa llvmpipe), on the same episodes. ∗ the two renderers differ, McNemar p<0.05 .
Policy
gain (points)
95% CI, tasks
95% CI, tasks within suite
tasks up/down/tied
sign test p
Wilcoxon p
OpenVLA-OFT
+15.1
[7.6,22.9]
[9.0,21.7]
26 / 5 / 9
1.9×10−4
4.5×10−4
π0.5
+18.6
[10.3,27.1]
[11.7,25.6]
23 / 9 / 8
0.020
0.0053
π0
+23.2
[17.1,29.7]
[17.5,28.9]
35 / 1 / 4
1.1×10−9
3.2×10−7
OpenVLA
+22.8
[17.7,28.4]
[18.4,27.7]
38 / 1 / 1
1.5×10−10
8.4×10−8
Appendix
Table S19 : Task-level resampling of the main result. Gain of Homeward over the frozen policy on all 1,599 LIBERO-Plus camera variants, with bootstrap intervals that resample the 40 tasks, and the number of tasks whose success count rises, falls or is unchanged.
Comparison (difference in points)
diff.
95% CI, tasks
Homeward minus AnyCamVLA, LIBERO-Plus camera variants
OpenVLA-OFT
−3.1 / −1.9
[−11.3,4.8] / [−9.2,5.3]
π0.5
+9.9 / +12.9
[3.5,16.8] / [6.4,19.7]
π0
+3.8 / +8.1
[0.3,7.5] / [4.7,12.5]
OpenVLA
+0.4 / +12.3
[−3.2,4.1] / [8.1,16.7]
Homeward minus AnyCamVLA (true cameras), orbit protocol on Long
Appendix
Table S20 : Task-level intervals for the other comparisons. Differences in success rate (points, from unrounded rates) with 95% bootstrap intervals that resample tasks (20,000 draws). Paired entries list true / estimated camera parameters.
Real-world robot deployment rarely maintains the training-stage camera setup, where cameras often experience repositioning or remounting depending on actual scenarios. Existing view-robust Vision-Language-Action (VLA) policies tolerate such camera variations only when the camera extrinsics are explicitly provided, making them fragile and hard to use especially when view robustness is critical. We argue that the policy should not be told where the camera is, but rather figure it out by itself. To this end, we introduce Camera-Centric VLA (CamVLA), a new VLA model that decouples manipulation controls from camera geometry by predicting (i) a camera-centric end-effector action expressed in the local camera frame, and (ii) a 6-DoF hand-eye matrix relating cameras to the robot base. A deterministic geometric transformation composes the two predictions into a robot base-frame action. This disentangles how I should move in pose-independent camera-centric action generation from where I am looking from in camera-perspective geometric grounding. The resulting policy is calibration-free, depth-free, and single-view, requiring only a single monocular RGB image as the visual observation and task instruction at deployment. Evaluations in both simulation and real-world robot data show that CamVLA consistently improves success rates across diverse unseen viewpoints. Project page: https://alibaba-damo-academy.github.io/CamVLA/.
Wenhao Li, Xueying Jiang, Quanhao Qian +4
1Nanyang Technological University · 2DAMO Academy, Alibaba Group · 3HuPan Lab +1
Vision-language-action (VLA) policies fine-tuned from a fixed scene camera can fail when the camera is moved, even when the task, objects, language, and robot state are unchanged. We study scene-camera viewpoint robustness using only a scene RGB image, language, and proprioception, without camera labels, extrinsics, depth, or point-cloud inputs. The wrist stream is masked throughout to prevent an unperturbed visual shortcut from confounding attribution to scene-camera variation. For flow-based VLAs, we propose to regularize the action-flow velocity field, the quantity directly integrated to generate continuous action chunks. We construct action-equivalent view pairs by resetting original LIBERO demonstrations to the same MuJoCo state and rendering nominal and perturbed scene-camera views. Both views are supervised by flow matching, while a cross-view loss encourages their predicted action-flow velocities to agree at the same sampled flow coordinates. On the LIBERO-Plus camera-perturbation track, our method reaches 87.2±0.4% (4,797 rollouts per seed across 3 training seeds), +7.4pp over flow-matching-only training on the same paired data (79.8±0.8%, also 3 seeds) and +12.5pp over naive mixed-camera SFT, while maintaining nominal-camera ID performance (95.0±0.8%; same-data FM-only: 95.0±4.3%). A shuffled-pair control collapses to 25.8%, showing that the gain depends on action-equivalent pairing. On a real robot, we evaluate three tabletop tasks with 10 rollouts per task and camera placement; held-out-camera success improves from 53.3% to 74.4% under the same single-scene-RGB inference interface.
Bingqi Huang, Bingchuan Wei, Xuan Wang +2
Tsinghua University · Informatics Institute, University of Amsterdam & Faculty of Science, Vrije Universiteit Amsterdam
This paper proposes a lightweight, plug-and-play framework that improves robustness to viewpoint shifts in Vision-Language-Action (VLA) policies without policy retraining. To our knowledge, this is the first approach to directly leverage 3D Gaussian-based novel-view synthesis for observation-space adaptation in VLA policies. Current VLA performance relies on the implicit assumption that training and deployment camera configurations are identical. Our experiments show that even a small displacement of the camera mount can reduce the success rate on the LIBERO benchmark from about 90% to about 10% in the worst case. Prior approaches, such as large-scale fine-tuning or generative data augmentation, are computationally expensive and risk catastrophic forgetting. To address this, viewpoint shifts are reformulated as a localized novel-view synthesis problem. Under a Locality assumption, that camera perturbations remain within a small bounded region relative to the workspace, viewpoint normalization reduces to a scene- and policy-independent disocclusion task. Our work implements this idea with a 4M-parameter 3D-Gaussian canonicalizer prepended to a frozen VLA policy. Without modifying policy weights, GS-VLA improves performance across three orthogonal axes: (1) Policy architectures, (2) Unseen task suites, and (3) Perturbation scales. These results show that a lightweight visual module can recover a large fraction of the performance lost under viewpoint shift, without policy retraining.
Yechan Park, HyunJin Kim
Dankook University Yongin, Republic of Korea · Dept. of EEE, Dankook University Yongin, Republic of Korea