RoboMamba
Mamba state-space VLA for efficient robot reasoning and manipulation — 3x faster inference than transformer VLAs.
About
RoboMamba (NeurIPS 2024, arXiv 2406.04339) swaps the transformer backbone for Mamba, the state-space model with linear-time inference. A vision encoder is co-trained with the Mamba language model for visual common sense and robot reasoning, then a tiny policy head is added for SE(3) pose prediction. The efficiency claim is the headline: once the model can reason, it picks up manipulation skills with just 0.1% of parameters fine-tuned, and inference runs about 3x faster than comparable VLA models. It shows strong results on reasoning benchmarks and on pose prediction in sim and real-world tests — all author-reported. Code is public at github.com/lmzpai/roboMamba, with checkpoints shared through the repo's test branch; the training code itself is only available by emailing the authors. Community uptake looks thin, which is the main risk: interesting architecture, limited independent mileage.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2406.04339): 3x inference speedup vs VLA baselines; reasoning and pose-prediction results (author-reported)
- No independent benchmarks found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER metrics, metrics-v1). Evidence: NeurIPS 2024 paper (arXiv 2406.04339) with sim and real-world pose-prediction evaluation; public code repo with checkpoints on the test branch. Strengths: novel Mamba-for-robotics recipe, strong efficiency claims. Docked for: (1) training code gated behind email request and thin community uptake; (2) all results author-reported; (3) license not confirmed. License: unverified — check repo before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · NeurIPS 2024 paper arXiv 2406.04339 (Jun 2024); Mamba state-space backbone + SE(3) pose prediction head
- · 0.1% fine-tune params, ~3x inference speedup vs transformer VLAs (author-reported)
- · Checkpoints on repo test branch; thin community uptake so far