TinyAstral-SFT

Paper: arXiv:2610.41216

Overview video (85 s): project homepage.

This is the supervised fine-tuning set used to train the TinyAstral students (4B / 2B / 0.8B). Each sample is one decision of an agent working inside the TinyAstral harness $H^*$. The input is the system prompt, the task, and the interaction so far (tool calls, tool results and camera images). The target is the next assistant turn: optional short reasoning followed by one tool call.

• 83,753 training samples and 4,443 validation samples.
• 123,836 images (4.2 GB), all with a .jpg extension: the RoboTwin, LIBERO and real-robot frames are JPEG, and the 10,726 Guava frames are PNG-encoded (standard image loaders read both). The JSONL files are 0.74 GB.
• Format: ms-swift / OpenAI-style messages with an images list. <image> placeholders in user turns are filled from images in order. Image paths are relative to the dataset root.
• Training uses loss on the final assistant turn only. Earlier assistant turns are context.
• Images are packed into 10 uncompressed tar files under images/ (one per source, e.g. images/sim_robotwin.tar, images/real_droid.tar) to keep the repository's file count low. Extract them in the dataset root to restore the relative paths used in the JSONL files:

  for t in images/*.tar; do tar -xf "$t"; done   # creates images/sim/... and images/real/...
  

• SHA256SUMS lists the checksums of the JSONL and tar files.

{"id": "...", "messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "Task: ..."},
  {"role": "assistant", "content": "<tool_call>...</tool_call>"}, {"role": "user", "content": "<image>Result of grasp: ..."},
  ..., {"role": "assistant", "content": "<think>...</think>\n\n<tool_call>{...}</tool_call>"}],
 "images": ["images/sim/robotwin/attempt0/<task>/<seed>/000_head.jpg", "..."]}

Composition

Subset	What it is	Train	Val
RoboTwin main	Frozen teacher in $H^*$, training seeds of the 50 RoboTwin 2.0 tasks	13,857	747
RoboTwin recovery	Branches from perturbed replays (object slips or is moved before the next grasp)	3,471	198
Guava main	7 seen Guava tasks	9,054	496
Guava recovery	Branches (missed grasp, short push, displaced object)	2,048	125
LIBERO main	LIBERO-90	11,159	1,329
LIBERO recovery	Branches	2,288	280
Real: verifier	Judge whether a sub-goal succeeded from real-robot frames	27,918	846
Real: next call	Next tool call for a real-robot episode, with reasoning	9,771	296
Real: planner	Sub-goal plan for a real-robot episode	4,187	126
Total		83,753	4,443

Teacher episodes run at temperature 0. A training seed on which the teacher failed was retried up to twice at temperature 0.7 (attempt1, attempt2 in image paths). Real-robot samples convert frames and annotations from public datasets into the harness format. Their reasoning was written by the teacher in hindsight, given the reference action, and filtered.

Real-robot sources (training samples / images)

Source	Verifier	Next call	Planner	Train total	Val	Images	License
AgiBot World	3,782	4,061	1,059	8,902	261	12,750	CC BY-NC-SA 4.0
RH20T	2,411	-	-	2,411	81	6,194	CC BY-SA 4.0 (scenes 1-5) / CC BY-NC 4.0 (scenes 6-10)
DROID (success episodes)	6,767	1,065	885	8,717	267	18,187	CC BY 4.0
DROID (failure episodes)	1,852	39	-	1,891	46	2,354	CC BY 4.0
RT-1 / Fractal (Open X-Embodiment)	5,442	2,052	939	8,433	257	9,222	CC BY 4.0
Bridge V2	4,760	1,670	1,044	7,474	204	7,884	CC BY 4.0
Berkeley UR5 (Open X-Embodiment)	2,904	884	260	4,048	152	5,275	CC BY 4.0

Simulated images: RoboTwin 39,048, Guava 10,726, LIBERO 12,196.

License

TinyAstral-SFT is released under CC BY-NC-SA 4.0 (non-commercial, share-alike), as one dataset. Some sources are non-commercial (AgiBot World, part of RH20T) and AgiBot World is share-alike, so the combined set carries CC BY-NC-SA 4.0.

Attribution. The dataset contains material derived from the following sources; each keeps its own license:

Source	Use in TinyAstral-SFT	License
DROID	real-robot frames and annotations (success and failure episodes)	CC BY 4.0
RT-1 / Fractal (Open X-Embodiment)	real-robot frames and annotations	CC BY 4.0
Bridge V2	real-robot frames and annotations	CC BY 4.0
Berkeley UR5 (Open X-Embodiment)	real-robot frames and annotations	CC BY 4.0
AgiBot World	real-robot frames and annotations	CC BY-NC-SA 4.0
RH20T	real-robot frames and annotations	CC BY-SA 4.0 (scenes 1-5) / CC BY-NC 4.0 (scenes 6-10)
RoboTwin 2.0	simulator; rendered frames from teacher episodes	MIT
Guava	simulator; rendered frames from teacher episodes	see upstream repository
LIBERO / LIBERO-PRO	simulator; rendered frames from teacher episodes (LIBERO-90)	MIT (LIBERO); see upstream for LIBERO-PRO

Teacher reasoning and tool calls were generated with Qwen3.8-Flash-Next and are also subject to its license. The students trained on this data (TinyAstral weights) are released under the same CC BY-NC-SA 4.0 terms.

Intended use

Training and analysing small decision models for tool-based embodied agents. Samples assume the TinyAstral harness prompt and tool interface.

Citation

@article{gong2026tinyastral,
  title         = {TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents},
  author        = {Gong, Kehong and Zheng, Xiaoxu and Hou, Guanli and Lee, Seunghun and Guo, Mengqi and Tan, Runjia and Wen, Zhengyu and Wang, Qi and He, Xiaoyu and Zhang, Mingyuan and Mi, Michael Bi},
  journal       = {arXiv preprint arXiv:2610.41216},
  year          = {2026},
  eprint        = {2610.41216},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.41216}
}