Nudge Before You Push: Physics-Aware Navigation via Tactile Probing
Authors: Xianyao Li, Fang Xu, Ruitong Tian, Bowen Sun, Xiao Hu, Yang Ye, Jing Du
Organizations: Department of Civil and Coastal Engineering, University of Florida · Department of Civil and Environmental Engineering, Northeastern University
Visually identical containers can conceal loads that require different handling decisions. We present TANav, which uses a brief nudge to measure push resistance for navigation under a site-defined handling boundary. TacPhys reads the force sequence, with optional RGB-D and kinematics, into a mass estimate for push authorization. A repeated-patrol planner weighs probe and route costs, requests a second contact when useful, and reuses observations across visits. In simulation, TacPhys approaches a resistance-only Bayes reference and reduces missed pushes from 28.7% to 5.5% relative to peak-force thresholding at comparable low-risk operating points. In repeated-patrol simulation, TANav recovers 90% of the oracle's path saving, more than halves human interventions relative to always-detour, and reduces boundary violations from 4.3% to 2.9% relative to RGB-D-only probing. On a quadruped manipulator with a Hall-array fingertip, offline zero-shot MAE is 0.28-1.07 kg on containers up to 2.82 kg. Force-rise calibration at 3 kg gives 93.5% pooled offline accuracy (86.4% on non-cube episodes); a separate raw-peak rule gives 15 of 20 correct online decisions on unseen boxes.
Figures & tables
Figure 1: Tactile probing for push-or-detour navigation. TANav separates contact interpretation, push eligibility, and route choice. In simulation, TacPhys reads nudge resistance into mass; the site boundary determines eligibility, and the patrol planner weighs probe and route costs, requests additional contact, and reuses cached observations. Hardware tests assess offline transfer and calibration; calibrated raw-peak decisions run online.
Figure 2: Platform and collection setup. Top: simulated and real B1+Z1 with the fingertip and head-mounted RGB-D camera. Bottom: a hardware contact and representative objects from the real-world dataset.
Input
Nudges
MAE (kg)
Acc. (%)
RGB-D
1
0.963±0.019
92.2
Tactile
1
0.757±0.007
94.5
Tactile + RGB-D
1
0.729±0.006
94.7
RGB-D
2
0.887±0.037
92.9
Tactile
2
0.665±0.008
95.3
Tactile + RGB-D
2
0.652±0.006
95.4
Table 1: Simulation estimation and decisions. MAE is mean ± standard deviation over five seeds; decision accuracy uses the common orthogonal test cohort.
Reader
Acc.
Test risk
Test missed
TacPhys , tactile + RGB-D + kin.
95.5
3.3
5.5
TacPhys , tactile only
95.3
3.2
6.1
RGB-D only
92.9
3.5
12.3
Ideal resistance, μmg
95.5
2.0
8.6
Peak force, first nudge
85.3
2.5
32.3
Mean peak force
86.3
2.6
28.7
Table 2: Decision baselines on 3,618 test episodes. Accuracy uses the default protocol; the final columns use a cutoff selected at a 3% validation risk target. For these columns, feature classifiers are also fitted on validation; test stays held out. All entries are percentages.
Encoder
Params
MAE (kg)
Acc.
Missed
Transformer, 6 layers
2.67M
0.694±0.007
95.1%
7.3%
BiGRU, 2 layers
0.33M
0.724±0.014
95.1%
7.3%
1-D CNN, 4 blocks
1.48M
0.733±0.013
95.0%
8.3%
Mean-peak threshold
–
–
86.3%
27.8%
Table 3: Controlled tactile-only encoders on the same 3,618 test episodes. MAE averages three seeds. Missed pushes use a diagnostic 3% test-risk sweep. Parameters count only the temporal encoder. All three lie near the resistance-only reference in Sec. 4.2 .
Figure 3: CoM sensitivity to force gain and contact onset (Sec. 4.2 ). (a) CoM accuracy and mass error under per-nudge force-gain variation. (b) CoM transfer from impulsive contact to a creep–rest–ramp protocol, before and after fine-tuning.
Figure 4: Hardware force readouts. Faint points are episodes; outlined points are object medians. The tested optical force regressor compresses peaks across loads. Hall-array readings track mass more closely. The dashed line marks 3 kg.
Instance
kg
ZS
Aff.
Rise
Acc.
Agree
Cube
1.69
0.45
0.43
5.3
96%
163/165
Box E1
1.76
0.28
0.49
8.1
83%
2/2
Box A1
2.46
0.71
0.57
6.9
100%
4/4
Box D1
2.82
1.07
0.61
8.4
100%
3/3
Waste bin
3.19
1.27
0.69
13.5
67%
4/4
Box D2
4.58
2.54
1.96
12.8
50%
1/1
Table 4: Hall-array results with each mass instance held out. ZS and Aff. are zero-shot and affine-calibrated TacPhys MAE (kg) over the 220 orthogonal pairs. Rise is the mean force increment (N); Acc. is the calibrated-boundary accuracy over all 262 pairs, and Agree gives correct / accepted episodes under offline agreement. Same-letter boxes share a container.
Policy
Saving
Interv.
Risk
Failed
Always detour
0%
0.26
0%
0
Oracle (true mass)
6.8%
0.11
0%
0
Try-push
5.6%
0.11
55.5%
2.6
Appearance prior
5.1%
0.11
55.5%
5.3
RGB-D during nudge
5.8%
0.12
4.3%
0.39
TANav , one nudge
6.1%
0.12
2.9%
0.33
Table 5: Repeated patrol over 300 layouts and ten cycles, with five seeds and a common planner at 3 kg. Saving is at cycle ten relative to always-detour; interventions are per patrol and failures per scene. Agreement variants are defined in Sec. 3.3 .
Figure 5: Repeated-patrol simulation. (a) Cumulative path saving at 3 kg. (b) Human interventions per patrol. (c) Path saving (solid) and boundary violations (dashed) as the authorization boundary varies. All policies share the planner; cached estimates are reused across cycles.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Example layout from the repeated-patrol simulation (Sec. 4.4 ). The nominal route passes through doorways containing blockers; orange paths show detours. Green and red markers distinguish loads at or below the 3 kg boundary from heavier loads.
School of Artificial Intelligence and Robotics, Hunan University, China · National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China · College of Semiconductors (College of Integrated Circuits), Hunan University, Changsha, China