Research map / Vision-language and multimodal / Robot action and navigation
RoboMamba: Efficient Vision-Language-Action Model for Robotic Reasoning and Manipulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces RoboMamba, an efficient VLA model pairing a CLIP vision encoder with the Mamba 2.8B/1.4B LLM, plus a tiny frozen-backbone policy head for end-effector pose prediction.
Problem
Existing MLLM-based robotic VLA models have limited robotic reasoning and are computationally expensive to fine-tune and run due to attention-based LLMs.
Approach
Replace the attention LLM with the linear-complexity Mamba SSM in a VLA, train general and robotic reasoning via alignment pre-training and co-training, then learn manipulation by freezing the whole model and training only a small MLP policy head (3.7M params, ~0.1%) on Mamba's output tokens.
Key results
- On RoboVQA, RoboMamba (2.7B, 224px) reaches 42.8 BLEU-4 and 62.7 BLEU-1, above fine-tuned TinyLLaVA (29.6 BLEU-4).
- On SAPIEN manipulation, RoboMamba averages 0.63 on seen and 0.53 on unseen categories, versus ManipLLM at 0.56 and 0.51.
- The fine-tuned policy head is 3.7M parameters (0.1% of the model), and inference is reported as about 3x to 7x faster than prior VLA models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mamba-2.7B instead of RWKV-3B (ablation) | RoboMamba reasoning model · General and robotic-related reasoning benchmarks | reasoning performance | Helped |
| Add RoboVQA robotic instruction data during co-training (ablation) | RoboMamba with LLaVA 1.5 instruction data · RoboVQA validation | BLEU-4: 26.5 -> 42.8 | Helped |
| Add RoboVQA robotic instruction data during co-training (ablation) | RoboMamba with LLaVA 1.5 instruction data · POPE | score: 85.6 -> 86.3 | Helped |
| Add RoboVQA robotic instruction data during co-training (ablation) | RoboMamba with ShareGPT4V-SFT · GQA | score: 64.3 -> 62.1 | Hurt |
| Add RoboVQA robotic instruction data during co-training (ablation) | RoboMamba with ShareGPT4V-SFT · RoboVQA validation | BLEU-4: 26.7 -> 42.5 | Helped |
| Add RoboVQA robotic instruction data during co-training (ablation) | RoboMamba with LLaVA-Next · RoboVQA validation | BLEU-4: 25.1 -> 43.0 | Helped |
| Replace LLaVA 1.5 instructions with ShareGPT4V-SFT (ablation) | RoboMamba without RoboVQA co-training · GQA | score: 65.3 -> 64.3 | No clear effect |
| Replace LLaVA 1.5 instructions with LLaVA-Next (ablation) | RoboMamba without RoboVQA co-training · RoboVQA validation | BLEU-4: 26.5 -> 25.1 | No clear effect |
| Replace the CLIP encoder with XCiT (ablation) | RoboMamba · RoboVQA validation | BLEU-4: 35.2 -> 42.8 | No clear effect |
| Change CLIP image resolution from 224 x 224 to 336 x 336 (ablation) | RoboMamba · RoboVQA validation | BLEU-4: 35.2 -> 41.8 | No clear effect |
| Replace CLIP with SigLIP and use 384 x 384 resolution (ablation) | RoboMamba · RoboVQA validation | BLEU-4: 35.2 -> 40.6 | No clear effect |
| Use one shared MLP head instead of separate position and direction heads (ablation) | RoboMamba manipulation policy · SAPIEN manipulation, seen | accuracy: 63.7% -> 62.1% | No clear effect |
Surprising findings
- Policy heads of very different sizes (1.8M MLP x1 vs 45.2M SSM+MLP) give similar manipulation accuracy, suggesting reasoning quality matters more than head capacity.
- Replacing CLIP with XCiT, an encoder without cross-modal alignment, gives the same reasoning scores as the main model (63.3 OKVQA, 42.8 BLEU-4), which the authors say is due to a mislabeled encoder in the table.
Limitations
- Reliance on a 2.7B LLM limits performance on complex reasoning relative to 7B/13B MLLMs; several benchmarks are below LLaVA1.5 and SPHINX.
- Manipulation is evaluated open-loop in simulation with 2D contact-point prediction, and real-world results are qualitative.
- Reasoning failures occur on complex tasks, such as misreading the task objective or target object.
Experimental setup
Datasets: LLaVA-LCS 558K LLaVA-v1.5 655K ShareGPT4V-SFT LLaVA-Next RoboVQA PartNet-Mobility SAPIEN simulation data
Benchmarks: OKVQA VQAv2 GQA VizWiz POPE MME MMBench MM-Vet RoboVQA SAPIEN manipulation benchmark
Metrics: accuracy BLEU-1 BLEU-2 BLEU-3 BLEU-4 success rate inference speed
Baselines: UMPNet FlowBot3D RoboFlamingo ManipLLM LLaMA-AdapterV2 TinyLLaVA RWKV-3B
Builds on: Mamba LLaVA ManipLLM RoboFlamingo Cobra
Closest papers by meaning
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.90)
- Vision-Language Foundation Models as Effective Robot Imitators (ICLR 2024 · similarity 0.90)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.90)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.