Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them

Carlota Parés-Morlans Nils Kuhn Y. Isabel Liu Alberta Longhini Jeannette Bohg
Stanford University
Paper (PDF) arXiv Code
0:00 / 0:00
Abstract

We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the factors contributing to these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures occur when policies struggle to produce the small, accurate corrective actions required during contact and force failures occur when policies fail to respond to contact forces. To address these failures, we propose FACT. During post-training, FACT increases low-noise sampling to improve fine action prediction and incorporates structured force conditioning to guide action generation using contact state and force history. FACT achieves an average success rate of 66% across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts.

Diagnosis

Compared to general manipulation tasks, contact-rich tasks require both fine-grained correction and reasoning about contact forces. Below, we illustrate these requirements through plug insertion, contrasting a successful execution with two distinct failure modes: precision failure and force failure.

Success

The plug approaches, aligns with the socket, enters and fully seats.

Precision failure

The policy fails to align the plug with the socket and stalls at the rim.

Force failure

The policy aligns the plug with the socket, but then fails to detect when the plug is fully seated.

We examine precision failures in terms of action magnitude and the distribution of training samples across noise levels. Contact sparsity, temporal structure, and sensitivity to contact state motivate our force-injection design.

Precision failures
  • Action scale during contact. In free space, action deltas are large and variable. During contact they decrease to near zero, and this low-delta regime must be reproduced precisely at the moment visual feedback is least informative.
  • Low-noise sampling during post-training. The Beta schedules used by π₀ [1], π₀.₅ [2], and SmolVLA [4] allocate relatively few sampled timesteps to τ < 0.2. This imbalance may hinder the acquisition of fine-grained corrections needed for reliable contact-rich behavior.
Force failures
  • Contact sparsity. Force signals are approximately zero over free-space timesteps and become informative only during short contact intervals. This imbalance makes naive concatenation prone to underusing informative contact readings.
  • Temporal structure. The instantaneous measurement encodes the current interaction state. Recent history encodes the dynamics that led to it, so omitting either timescale discards task-relevant contact information.
  • Sensitivity modulation. The influence of force should depend on contact state. In free space, force readings should be effectively ignored, whereas at contact even small deviations should trigger corrective actions.
Approach

Each failure mode is addressed with a targeted mechanism, and together they form FACT, Force-Aware Contact-rich manipulation via Timestep modulation. LN preserves the architecture, while force injection adds conditioning modules. Both require no additional data and add negligible inference-time overhead.

Precision · Logit-Normal noise schedule

We adopt Logit-Normal (LN), an established sampling distribution for rectified flow training [7], to increase exposure to low noise levels during post-training. We use a location shift m=1.5 and scale s=1, shifting the distribution toward lower noise levels compared with m=0. LN allocates approximately 6× more sampling probability to τ < 0.2 than the default Beta schedule, while maintaining support across the noise range. This concentrates post-training on reconstructing actions from small perturbations, with the aim of improving precise corrective actions during contact. LN requires no changes to the model architecture, no additional training data, and no extra parameters.

Beta
Logit-Normal
LN allocates 6.0× the sampling probability of Beta below τ = 0.20
Sampling probability density over noise level τ. Drag to explore the schedule.
Force · Time-aware force injection

Three design choices, one per force-signal property. The added force-conditioning modules use approximately 2.2M parameters.

Contact state injected into AdaRMSNorm layers

For the full ablation study and hyperparameter details, please refer to the paper.

Results

FACT achieves the highest average success rate across the five contact-rich tasks, with gains on both precision- and force-critical tasks.

39%
π₀.₅ [2]
38%
TA-VLA [6]
41%
ForceVLA [5]
56%
π₀.₅ + LN (ours)
66%
FACT (ours)

Average success rate (%) across all five tasks.

Method PlugUSBButtonBoardKeyAll ↑
π₀.₅ [2]30.037.512.510015.039.0
TA-VLA [6]30.025.020.097.515.037.5
ForceVLA [5]32.537.512.577.542.540.5
π₀.₅ + LN (ours)50.047.557.587.537.556.0
FACT (ours)57.547.575.090.060.066.0

Success rate (%) per task, evaluated over 40 rollouts per method per task. All methods are fine-tuned from π₀.₅ with LoRA using 100 demonstrations per task for 20K steps. Tasks span precision-critical (plug, USB) and force-critical (button, board, key) regimes.

References
  1. K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024.
  2. P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al. π0.5: A vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025.
  3. J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, et al. GR00T N1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025.
  4. M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. SmolVLA: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025.
  5. J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y. Song, P. Cai, et al. ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation. arXiv preprint arXiv:2505.22159, 2025.
  6. Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. TA-VLA: Elucidating the design space of torque-aware vision-language-action models. In 9th Conference on Robot Learning (CoRL), 2025.
  7. P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In 41st International Conference on Machine Learning (ICML), 2024.
BibTeX
Cite this work
@article{paresmorlans2026fact,
  title={Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them},
  author={Parés-Morlans, Carlota and Kuhn, Nils and Liu, Y. Isabel and Longhini, Alberta and Bohg, Jeannette},
  journal={arXiv preprint arXiv:2608.01402},
  year={2026}
}
Acknowledgements

This work was supported in part by Agile Robotics. Carlota Parés-Morlans is supported by a graduate fellowship from Knight-Hennessy Scholars at Stanford University. Nils Kuhn is supported by scholarships from the Friedrich Ebert Foundation and the German Academic Exchange Service (DAAD). Alberta Longhini is supported by a Wallenberg–Bienenstock Postdoctoral Fellowship.

We thank Hila Chefer for helpful discussions, and Michelle Yi for discussions on environment setup, data collection, and camera footage for figures.