Model card
- Made by
- Galaxea
- Size
- 2B parameters
- Published
- 12 Aug 2026
More about the model
- Publication
- arXiv 2608.11739
- Weights licence
- G0.5 Community License: non-commercial use only (research, education, evaluation) source
- Actions per chunk
- 32 source
- Smallest GPU
- About 12 GB of free GPU memory for the DROID checkpoint source
- Architecture
- Qwen3.5 2B vision-language model that writes reasoning and action tokens in one stream; a learned action codec decodes the action tokens source
- Training data
- Base model: robot demonstrations from 14 embodiments, and web and embodied VQA data; this checkpoint is post-trained on DROID source
Progress through an episode
Each band shows how many episodes of G0.5 were in that band at each second. On a task of one step, an episode is in the band of the furthest stage it reached. On a task of several steps, an episode is in the band its credit has reached, and each band starts at its stage's credit for one step. The dashed line shows the mean Reliability.
Episodes
Each episode of G0.5 on DROID: Everyday tasks, the newest first.
Progress through an episode
Each band shows how many episodes of G0.5 had succeeded, partly succeeded (some subtask score) or had no success at each second of simulated time. An episode that has ended stays in its band. The dashed line shows the mean Reliability, the mean subtask score.
Episodes
Each episode of G0.5 on RoboLab-120, the newest first.
Progress through an episode
Each band shows how many episodes of G0.5 had succeeded or had no success at each second of simulated time. An episode that has ended stays in its band. The dashed line shows the mean Reliability, the share of the episodes that had succeeded.
Episodes
Each episode of G0.5 on MolmoSpaces-320, the newest first.