BodyVLA: Embedding Morphology into Transformers for Cross-Robot Policy Learning

An embodiment-aware transformer policy that injects morphology.

MERL Researchers: Toshiaki Koike-Akino, Kei Suzuki, Jing Liu, Ye Wang, Matthew Brand, Chiori Hori.

---

Paper Appendix_along_with_Main_Paper.pdf

Video Supplemental_Material_Video.mp4

---

---

Abstract

Cross-robot policy learning—training a single policy to perform well across multiple embodiments—remains a central challenge in robot learning. Transformer-based policies, such as vision-language-action (VLA) models, are typically embodiment-agnostic and must infer kinematic structure purely from observations, which can reduce robustness across embodiments and even limit performance within a single embodiment. We propose an embodiment-aware transformer policy that injects morphology via three mechanisms: (1) kinematic tokens that factorize actions across joints and compress time through per-joint temporal chunking; (2) a topology-aware attention bias that encodes kinematic topology as an inductive bias in self-attention, encouraging message passing along kinematic edges; and (3) joint-attribute conditioning that augments topology with per-joint descriptors to capture semantics beyond connectivity. Across a range of embodiments, this structured integration consistently improves performance over a vanilla p0.5 VLA baseline, indicating improved robustness both within an embodiment and across embodiments.

Method

We propose an embodiment-aware transformer policy through a structured morphology interface consisting of three components:

Results

DROID rollout videos: π0.5 Baseline vs. Ours

DROID Task 1

π0.5 Baseline: ✅ Success

Ours: ❌ Failure

DROID Task 2

π0.5 Baseline: ❌ Failure

Ours: ✅ Success

DROID Task 3

π0.5 Baseline: ❌ Failure

Ours: ✅ Success

SO101 and Unitree G1 Simulation Rollouts

DROID—SO101 Joint Training and Real-World Experiment

We evaluate DROID—SO101 joint-trained policies on a real SO101 robot.

Method	Config. 1	Config. 2	Config. 3	Total
p0.5 Baseline	0	3	1	4
Ours	11	9	13	33
Improvement	+11	+6	+12	+29

SO101 rollout videos

10 trials - Baseline 0 oranges - Ours 11 oranges

π0.5 Baseline ❌ 0 oranges placed

Ours ✅ 2 oranges placed

10 trials - Baseline 3 oranges - Ours 9 oranges

π0.5 Baseline ❌ 0 oranges placed

Ours ✅ 2 oranges placed

10 trials - Baseline 1 orange - Ours 13 oranges

π0.5 Baseline Δ 1 orange placed

Ours ✅ 3 oranges placed

Ablation Studies

To validate the effectiveness of our design choices, we study whether performance depends on (i) the kinematic token chunk size G, (ii) auxiliary kinematic tokens (AKT), and (iii) bias initialization for the Soft-Mask variant.

Citation

•  Suzuki, K., Liu, J., Wang, Y., Hori, C., Brand, M., Romeres, D., Koike-Akino, T., "Embedding Morphology into Transformers for Cross-Robot Policy Learning", arXiv, February 2026.
  BibTeX arXiv

  • @article{Suzuki2026feb,
  • author = {Suzuki, Kei and Liu, Jing and Wang, Ye and Hori, Chiori and Brand, Matthew and Romeres, Diego and Koike-Akino, Toshiaki},
  • title = {{Embedding Morphology into Transformers for Cross-Robot Policy Learning}},
  • booktitle = {arXiv},
  • year = 2026,
  • month = feb,
  • url = {https://arxiv.org/abs/2603.00182}
  • }

MERL Publications

•  Suzuki, K., Liu, J., Wang, Y., Hori, C., Brand, M., Romeres, D., Koike-Akino, T., "Embedding Morphology into Transformers for Cross-Robot Policy Learning", International Conference on Learning Representations (ICLR) Workshop, April 2026.
  BibTeX TR2026-048 PDF

  • @inproceedings{Suzuki2026apr,
  • author = {Suzuki, Kei and Liu, Jing and Wang, Ye and Hori, Chiori and Brand, Matthew and Romeres, Diego and Koike-Akino, Toshiaki},
  • title = {{Embedding Morphology into Transformers for Cross-Robot Policy Learning}},
  • booktitle = {International Conference on Learning Representations (ICLR) Workshop on Efficient Spatial Reasoning},
  • year = 2026,
  • month = apr,
  • url = {https://www.merl.com/publications/TR2026-048}
  • }