Skip to content
← Models
Open weightsByteDance Research / Tsinghua University/10 trustVerified
RoboFlamingo logo

RoboFlamingo

Turns the open-source OpenFlamingo VLM into a robot imitator with a small explicit policy head — strong on CALVIN.

Visit official page →

About

RoboFlamingo (November 2023, arXiv 2311.01378; ICLR 2024) is one of the earliest 'just fine-tune a VLM' robot policies. It takes OpenFlamingo, uses it for single-step vision-language comprehension, models the action history with a small explicit policy head on top, and fine-tunes the whole thing by imitation learning on language-conditioned manipulation data. The decomposition keeps inference cheap enough to train and run on a single GPU server — a deliberate contrast with the big-lab VLA recipes. On the CALVIN long-horizon manipulation benchmark it beat the state of the art at the time by a wide margin. Code is public at github.com/RoboFlamingo/RoboFlamingo. A follow-up, RoboFlamingo-Plus, adds depth fusion for another 10–20%. The CALVIN numbers are author-reported and simulation-based, so real-world transfer is untested in the paper.

Specs

Context window

Not applicable — image history + instruction in, actions out

Pricing

Free / open weights

Free quota

Fully open — no API

API access

No hosted API — public repo github.com/RoboFlamingo/RoboFlamingo

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Paper (arXiv 2311.01378): SOTA on CALVIN benchmark by a large margin (author-reported, simulation)
  • No independent replication found

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

roboticsrobot-manipulationvision-language-actionimitation-learning

Trust score — 7/10

/10 trustVerified

Scored 7/10 (TIER verified, metrics-v1). Evidence: ByteDance/Tsinghua paper (arXiv 2311.01378, ICLR 2024) with CALVIN benchmark results; public code repo. Strengths: early influential VLA design, cost-effective single-GPU recipe, open release. Docked for: (1) benchmark is simulation-only and author-reported; (2) repo looks quiet since release. License: unverified — check repo before commercial use.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · ByteDance/Tsinghua paper arXiv 2311.01378 (Nov 2023), ICLR 2024; OpenFlamingo + explicit policy head
  • · SOTA on CALVIN long-horizon manipulation benchmark at release (author-reported, simulation)
  • · Single-GPU-server trainable; follow-up RoboFlamingo-Plus adds depth fusion

Sources