Research project · 2026
PSR learns representations from multimodal sensorimotor signals by jointly forecasting future force/torque sequences, future joint-effort sequences, and terminal joint-state displacement. PSR-VLA integrates the resulting hierarchy into selected Action Expert layers through gated cross-attention.
Real-world task suite
PSR-VLA rollouts span coupled assembly, precise insertion, and surface contact, with 20 consecutive evaluation trials per task.
Coupled alignment and deformable insertion
Precise multi-contact alignment
Broader feasible contact regions
Evaluation protocol
Long-horizon stability tests
These stress tests reuse the PSR-VLA checkpoints from the main experiments without expanding the training sets or performing additional training. They are reported separately and are not included in the six-task success-rate averages.
442/450successful trials
98.2%success rate
9test positions
50trials per position
48/50successful trials
96.0%success rate
50consecutive trials
0retries or recovery
Additional evaluations
These supplementary experiments broaden the evaluation beyond the fixed six-task comparison and are reported separately from the success-rate averages.
Method
PSR receives compressed multiview visual tokens together with force/torque, joint-state, and joint-effort histories. A six-layer predictive Transformer learns an ordered hierarchy of sensorimotor representations that conditions selected layers of the Action Expert.
The first two Transformer layers preserve modality-specific cues, the middle three exchange information through bottleneck tokens, and the final layer refines the fused representation. A shared decoder reads all six retained layers and predicts future force/torque, future joint effort, and terminal joint-state displacement.
During action learning, the six PSR representations are aligned in depth order with Action Expert layers 3, 6, 9, 12, 15, and 18 through gated cross-attention. The prediction decoder is frozen and no prediction loss is applied in Stage 2; learnable gates regulate each residual sensorimotor update.
Quantitative evaluation
Success rates are measured over 20 consecutive trials per task. Tasks follow the hard-to-easy order in the experimental setup; Overall reports total successes over 120 trials.
Under the same backbone and evaluation protocol, PSR-VLA reaches 110/120 successful trials, compared with 74/120 for π0.5, 83/120 for ForceVLA-π0.5, and 87/120 for ForceVLA2-π0.5. These correspond to gains of 30.0, 22.5, and 19.2 percentage points, respectively. PSR-VLA achieves the highest success rate on five of the six tasks.
Across the tabulated run and three additional PSR-VLA training runs, performance is 109.5 ± 2.5 successes out of 120, or 91.3 ± 2.1% (mean ± s.d.). This stability analysis is PSR-VLA-only and is not a seed-matched comparison with the baselines.
The gains remain consistent from hard tasks (92.5% mean) to medium (85.0%) and easy tasks (97.5%). Three-Hole Plug Insertion is the only task not led by PSR-VLA, where ForceVLA2-π0.5 reaches 85% compared with 80% for PSR-VLA.
Matched qualitative comparison
Both policies begin the Three-Hole Plug task with closely matched plug-to-socket alignment. As contact loads build, π0.5 develops lateral drift and ends off-axis. PSR-VLA preserves axial alignment and directs the applied force vertically downward, completing the insertion.
Representative paired rollouts with comparable initial alignment. This example illustrates contact behavior and is not used to estimate the overall success rate.
Predictive representation
On held-out sequences, PSR reduces normalized MSE over a persistence baseline by 54.9% for future force/torque, 48.1% for future joint effort, and 87.5% for terminal joint-state displacement.
Ablation studies
Three controlled questions test predictive pretraining, ordered depth-wise representations, and multi-depth integration on Water-Hose Insertion and Two-Hole Plug Insertion. Each cell reports real-robot success together with normalized offline action error.
20 robot trials · 100 offline episodes per task
| Variant | Water-Hose Insertion | Two-Hole Plug Insertion | Average |
|---|---|---|---|
| w/o Stage-1 PretrainingNo predictive initialization | 14/2070% successMSE .005190L1 .026432 | 11/2055% successMSE .004378L1 .033470 | 62.5%mean success |
| PSR-VLAPrediction-pretrained model | 19/2095% successMSE .003378L1 .021413 | 19/2095% successMSE .003816L1 .035424 | 95.0%mean success |
| Variant | Water-Hose Insertion | Two-Hole Plug Insertion | Average |
|---|---|---|---|
| Shared Deepest RepresentationReuse z6 at all six injection sites | 14/2070% successMSE .004378L1 .031470 | 15/2075% successMSE .004066L1 .036335 | 72.5%mean success |
| Reversed AlignmentReverse representation-to-layer pairing | 13/2065% successMSE .004696L1 .029665 | 12/2060% successMSE .004807L1 .041436 | 62.5%mean success |
| PSR-VLAOrdered layer-to-depth pairing | 19/2095% successMSE .003378L1 .021413 | 19/2095% successMSE .003816L1 .035424 | 95.0%mean success |
| Variant | Water-Hose Insertion | Two-Hole Plug Insertion | Average |
|---|---|---|---|
| Output-Level ConditioningCondition only at the Action Expert output | 10/2050% successMSE .005969L1 .035745 | 11/2055% successMSE .004438L1 .041181 | 52.5%mean success |
| PSR-VLASix depth-aligned injection sites | 19/2095% successMSE .003378L1 .021413 | 19/2095% successMSE .003816L1 .035424 | 95.0%mean success |
Resources
Venue, paper, code, and model links will be updated when the public release information is confirmed.
@article{anonymous2026psrvla,
title = {PSR-VLA: Predictive Sensorimotor Representation Learning
for Contact-Rich Manipulation},
journal = {Under Review},
year = {2026}
}