Learned manipulation policies often rely on wrist-mounted cameras for close-range visual feedback that complements external views of the workspace. However, contact-rich manipulation requires establishing and maintaining contacts that can remain hidden even from wrist cameras. Tactile sensors mounted on fingers can reliably and directly measure contact, without needing visual access. In this work, we show that this feedback enables policies to recover the performance advantage of wrist cameras using scene and egocentric vision together with touch. Our approach incorporates two low-cost tactile arrays into a visuotactile ACT policy, without wrist images during training or deployment. We evaluate six real-world task families spanning placement, insertion, occluded retrieval, and sustained-contact tool use. On three contact-rich tasks, we independently vary wrist-camera and tactile availability while keeping the demonstrations, policy architecture, and optimization settings fixed. Adding wrist cameras improves success by 18.3-28.3 percentage points over vision without wrist images. Adding tactile sensing instead improves success by 31.7-60.0 percentage points and yields higher observed success than vision-only policies with wrist cameras on all three tasks, with significant advantages on two under Holm-corrected tests. These results support finger-mounted tactile sensing as an effective alternative to wrist cameras in the evaluated contact-rich settings.
Figures & tables
Figure 1: System and evaluation tasks. Two FlexiTac pads are mounted on the inner gripper surfaces and recorded alongside scene, egocentric, and wrist RGB observations. Panels 1(d) – 1(i) show tasks T1–T6, spanning placement, fine alignment, occluded retrieval, and sustained contact.
Task
Wrist
V
V+T
Δ
T1
Tight-clearance block
on
15/30 (50.0%)
19/30 (63.3%)
+13.3pp
T2a
Rigid block
on
15/25 (60.0%)
15/25 (60.0%)
0.0pp
T2b
Sponge
on
19/25 (76.0%)
22/25 (88.0%)
+12.0pp
T3
Straw insertion
on
26/60 (43.3%)
39/60 (65.0%)
+21.7pp
T4
Opaque-bag retrieval
off
13/60 (21.7%)
49/60 (81.7%)
+60.0pp
on
30/60 (50.0%)
43/60 (71.7%)
+21.7pp
Table 1: Vision-only vs. vision+tactile success. Each policy entry reports successes/rollouts (success rate). T4–T6 are evaluated both with and without wrist images. Δ is the observed V+T -minus- V difference in percentage points.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Customized station design
Task
Recorded
Successful
T1: Block placement
267
229
T2: Block/sponge dataset
225
208
T3: Straw insertion
220
209
T4: Bag retrieval
235
203
T5: Plug insertion
231
202
T6: Whiteboard wiping
218
207
Appendix
Table 2: Demonstration dataset sizes. Total recorded episodes and successful episodes before the episode-level training/validation split.
Setting
Value
Action space
7-D absolute joint positions
Action chunk
19 consecutive steps
KL weight
10.0
Batch size
48
Optimizer
AdamW
Learning rate
10−5
Appendix
Table 3: ACT training settings
Task
Initial conditions
Trials per position
Total
T1
Middle, top left, top right, bottom left, bottom right, far right in front of the box
5 each
30
T2a/b
Middle, top left, top right, bottom left, bottom right
5 each
25 each
T3
Middle, top left, top right, bottom left, bottom right
12 each
60
T4
Far left, left, middle, right, far right
10, 5, 5, 20, 20
60
T5
Left, right, middle, top, bottom
12 each
60
T6
Top, center, bottom
20 each
60
Appendix
Table 4: Evaluation allocation per observation condition.
Figure 3: Tactile gain by task and wrist-camera availability. Posterior mean and 90% credible interval for the difference in success probability between V+T and V .
Figure 4: Posterior success probability for every task, with compact letter displays. Beta posteriors under a uniform prior. Top row: the wrist-on tasks (T1–T3), two conditions each. Bottom row: the ablated tasks (T4–T6), four conditions each. Letters come from Holm-corrected pairwise Barnard tests computed within each panel; conditions sharing a letter could not be distinguished at α=0.05 . Letters are not comparable across panels. In T4 and T6 the two V+T conditions share a letter while the two V conditions do not—no significant wrist-on versus wrist-off difference was detected for V+T in these tasks.
Task
Initial condition
V
V+T
T1
Middle
4/5
5/5
Top left
2/5
3/5
Top right
0/5
2/5
Bottom left
1/5
0/5
Bottom right
5/5
5/5
Far right in front of the box
3/5
4/5
Appendix
Table 5: T1–T3 success counts by initial condition. Wrist images are enabled. Each entry is successes/rollouts.
Wrist off
Wrist on
Initial position
V
V+T
V
V+T
Far left
3/10
9/10
5/10
8/10
Left
4/5
5/5
5/5
4/5
Middle
2/5
4/5
5/5
5/5
Right
4/20
16/20
11/20
17/20
Far right
0/20
15/20
4/20
9/20
Appendix
Table 6: T4 retrieval success by initial position. Each entry is successes/rollouts. The aggregate V+T wrist-on decrease is concentrated in the far-right group; the other positions yield 34/40 successes in both V+T configurations.
Wrist off
Wrist on
Initial position
V
V+T
V
V+T
Left
0/12
8/12
3/12
7/12
Right
1/12
3/12
5/12
3/12
Middle
3/12
8/12
6/12
7/12
Top
4/12
4/12
4/12
6/12
Bottom
1/12
5/12
2/12
7/12
Appendix
Table 7: T5 insertion success by initial position. Each entry is successes/rollouts.
Wrist off
Wrist on
Initial position
n
V
V+T
V
V+T
Top
20
(5,4)
(17,13)
(8,7)
(18,15)
Center
20
(15,6)
(16,15)
(17,15)
(15,15)
Bottom
20
(3,3)
(15,15)
(6,6)
(9,9)
Total
60
(23,13)
(48,43)
(31,28)
(42,39)
Appendix
Table 8: T6 pickup and task success by initial position. Each policy entry is (P,S) , where P is the number of successful pickups and S is the number of successful complete trials. The trial count n applies to every policy in that row.
Department of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA · Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA