tanvirnwu/SUREFlow_IROS_2026

Accepted at IEEE/RSJ International Conference on Intelligent Robots & Systems (IROS) 2026, Pittsburgh, PA, USA

SUREFlow: State-space Uncertainty-aware REsidual Flow Matching for Robust Robot Manipulation (Paper)

Md Tanvir Islam, Sai Navaneet Peddapalli, Sangmoon Lee, Sangtae Ahn^*

Kyungpook National University, Daegu 41566, Republic of Korea | *Corresponding Author

---

SUREFlow Architecture

SUREFlow is a lightweight, Mamba-based vision-language-action model with 179M parameters for robot manipulation.

LIBERO Benchmark Results

Method	Venue	Spatial	Object	Goal	Long	Average
Octo [13]	RSS'24	78.9	85.7	84.6	51.1	75.1
QueST [16]	NeurIPS'24	89.0	90.0	88.4	87.0	88.6
MAIL [3]	CoRL'24	53.8	81.5	56.3	41.7	58.3
TraceVLA [17]	ICLR'25	84.6	85.2	75.1	54.1	74.8
SUREFlow (Ours)	IROS'26	94.8	91.0	93.8	90.2	92.5

LIBERO-PRO Benchmark Results

TABLE II

LIBERO-PRO model leaderboard showing normalized success rates under five perturbation types across four benchmarks.

Model	Goal Obj	Goal Pos	Goal Sem	Goal Task	Goal Env	Spatial Obj	Spatial Pos	Spatial Sem	Spatial Task	Spatial Env	LIBERO-10 Obj	LIBERO-10 Pos	LIBERO-10 Sem	LIBERO-10 Task	LIBERO-10 Env	Object Obj	Object Pos	Object Sem	Object Task	Object Env	Average SR ↑	Params ↓
OpenVLA [12]	0.96	0.00	0.98	0.00	0.98	0.97	0.00	0.97	0.00	0.89	0.81	0.00	0.96	0.00	0.85	0.98	0.00	0.98	0.00	0.00	0.52	7B
π0 [28]	0.94	0.00	0.93	0.00	0.39	0.95	0.00	0.97	0.00	0.60	0.79	0.00	0.82	0.00	0.27	0.94	0.00	0.90	0.00	0.29	0.44	3B
π0.5 [15]	0.97	0.38	0.97	0.00	0.46	0.97	0.20	0.97	0.01	0.46	0.92	0.08	0.93	0.01	0.46	0.98	0.17	0.96	0.01	0.73	0.53	3B
SUREFlow (Ours)	0.93	0.00	0.89	0.00	0.93	0.92	0.00	0.90	0.00	0.93	0.21	0.00	0.78	0.00	0.74	0.68	0.00	0.94	0.00	0.91	0.49	179.1M

SUREFlow

SUREFlow supports training on LIBERO suites and evaluating a trained checkpoint on either:

• the same vanilla LIBERO suite, or
• a LIBERO-PRO variant of that suite.

This is done by decoupling:

• train suite: dataset + training language embeddings
• eval suite: simulator benchmark + evaluation language embeddings

When --eval_suite is set, simulator benchmark becomes: <train_suite>_<eval_suite>

Examples:

• train_suite=libero_goal, eval_suite=object -> sim benchmark libero_goal_object

Training

Train on a vanilla LIBERO suite:

python run.py --train_suite libero_spatial

Supported --train_suite values: libero_object, libero_spatial, libero_goal, libero_90, libero_10

Evaluation with a checkpoint

1) Vanilla LIBERO evaluation

Evaluate a checkpoint on the same vanilla suite:

python run.py --train_suite libero_spatial --checkpoint_path /path/to/ckpt.pth

2) LIBERO-PRO evaluation on the same checkpoint

Evaluate the same checkpoint on a LIBERO-PRO suite:

python run.py --train_suite libero_spatial --eval_suite object --checkpoint_path /path/to/ckpt.pth

Eval suites (LIBERO-PRO suffixes)

Optional --eval_suite values: object, swap, lan, task, temp

In this mode:

• dataset benchmark remains libero_goal
• simulator benchmark is libero_goal_object
• evaluation embeddings are loaded from language_embeddings/libero_goal_object.pkl

Repository notes

The public package is SUREFlow. The original Mamba implementation is kept under SUREFlow/mamba/ so the backbone code remains easy to compare with the upstream block implementation.

Cite this Paper

If you find our work useful in your research, please consider citing our paper and star ✨✨ this repository. Thank you!

@article{islam2026sureflow,
  title={SUREFlow: State-space Uncertainty-aware REsidual Flow Matching for Robust Robot Manipulation},
  author={Islam, Md Tanvir and Peddapalli, Sai Navaneet and Lee, Sangmoon and Ahn, Sangtae},
  journal={arXiv preprint arXiv:2607.10504},
  year={2026}
}