BFM-Zero nominal tracking
GT, BFM-Zero, Larger BFM-Zero, and PRISM side by side.
A lightweight proprioceptive conditioner that learns polynomial proprioceptive feature mixtures for modern locomotion and vision-language-action policies.
PRISM adds a small learnable proprioceptive conditioner to strong robot-policy backbones, improving BFM-Zero tracking and SmolVLA LIBERO success over both the original models and larger-conditioner baselines.
Following the VIGOR-style layout, qualitative rollouts are visible before the detailed tables.
BFM videos are aligned to the 9.6M checkpoint comparison used in the locomotion table. LIBERO videos are same-episode comparisons selected from the official 80K eval50 logs, with aggregate metrics reported below.
PRISM targets proprioceptive context, not the visual backbone or attention mechanism.
Physical robot behavior depends on multiplicative interactions between joint state, velocity, contact, load, and task context. Standard proprioceptive encoders often compress these signals with shallow linear or MLP projections, which can make useful higher-order structure harder to discover.
PRISM adds a lightweight polynomial conditioning module over proprioception. Instead of replacing the base policy or hand-coding a fixed feature set, it uses a polynomially structured initialization and learns the resulting feature mixture end-to-end.
We validate this idea on two stronger backbones: BFM-Zero for humanoid locomotion and SmolVLA for LIBERO manipulation. Across both domains, PRISM improves over the standard backbone and a larger-capacity conditioner, suggesting the gain is not merely a parameter-count effect.
Main results use aligned checkpoints and official evaluation protocols where available.
At the aligned 9.6M checkpoint, PRISM reaches the lowest joint-position tracking error while adding only 0.556M inference parameters over BFM-Zero and using far fewer parameters than the larger baseline.
At 80K steps with official LIBERO multi-task eval50, PRISM improves average success over both SmolVLA and a larger-conditioner SmolVLA baseline.
BFM-Zero reports average joint-position tracking error between generated and reference motions. Lower is better.
| Setting | Method | Params | Joint Pos. Err. ↓ | Capacity Control |
|---|---|---|---|---|
| BFM-Zero | BFM-Zero | 32.109M | 1462.8 | -- |
| BFM-Zero | Larger BFM-Zero | 49.260M | 1456.1 | yes |
| BFM-Zero | PRISM | 32.665M | 1404.2 | yes |
Scores are averaged across nominal, low-friction, and payload/mass perturbation evaluations. At this early checkpoint, sparse success is less informative, so the tracking metric is used as the primary motion-quality indicator.
Official LIBERO multi-task eval50 at the 80K checkpoint. Higher success is better.
| Method | Total Params | Spatial | Object | Goal | Long | LIBERO Avg. ↑ |
|---|---|---|---|---|---|---|
| SmolVLA | 450.046M | 69.8 | 57.2 | 81.2 | 45.8 | 63.50 |
| Larger SmolVLA | 456.245M | 64.6 | 65.6 | 85.0 | 44.4 | 64.90 |
| PRISM | 451.925M | 70.0 | 57.4 | 85.4 | 53.4 | 66.55 |
PRISM adds 1.879M parameters over SmolVLA (+0.42% total) while using 4.321M fewer parameters than the larger-conditioner baseline.
The module makes a targeted architectural bet: proprioception benefits from a learnable polynomial inductive bias, not a fixed hand-engineered feature map.
PRISM starts from a polynomially structured conditioner, making low-order proprioceptive interactions easy for the policy to access early in training.
During training, the conditioner learns how much each dimension should behave linearly, quadratically, or as a mixed feature for the task.
The design avoids hard-coding a fixed state-plus-square representation; polynomial structure is used as a learnable inductive bias.
Placeholder citation for the project page.
@article{anonymous2026prism,
title = {PRISM: Polynomial Proprioceptive Conditioning for Generalist Robot Policies},
author = {Anonymous Authors},
journal = {},
year = {2026}
}