botware.club

Open AI models for robots

Hand-picked open models that run on robots and edge boards, checked against their source.

Vision

• Ultralytics YOLO
  The popular real-time YOLO family: detect, segment, classify and track objects, with exports to ONNX, TensorRT and Edge TPU.
• RT-DETR (R50)
  Real-time transformer detector trained on COCO's 80 everyday object classes.
• DETR ResNet-50
  The original end-to-end transformer object detector, trained on COCO.
• YOLOS tiny
  A very small vision-transformer detector for COCO objects.
• Grounding DINO tiny
  Finds any object you describe in words ("red screwdriver"), no training needed.
• OWLv2
  Detects objects from a text query or an example image.
• MobileNet V2
  Classic lightweight ImageNet classifier, a good starting point for small boards.
• ResNet-50
  Widely used ImageNet classifier and backbone.
• ViT base
  Vision Transformer classifier fine-tuned on ImageNet.
• SAM 2.1 tiny
  Segment anything in images and video from a click or a box.
• SegFormer B0
  Small scene segmentation model (floor, wall, person… 150 classes) for navigation.
• Depth Anything V2 small
  Depth map from a single ordinary camera image.
• DPT hybrid (MiDaS)
  Monocular relative depth estimation.
• ViTPose base
  Human body keypoints (17 joints) from person boxes.
• MediaPipe
  Ready-made on-device solutions: hands, face, body pose, objects, gestures.
• TrOCR printed
  Reads a line of printed text from an image crop.
• PaddleOCR
  Full OCR toolkit (find + read text) in 80+ languages, including Hindi, Kannada, Tamil and Telugu.
• EasyOCR
  Simple Python OCR for 80+ languages and scripts.

Vision-language

• SmolVLM 256M
  Tiny model that answers questions about images; small enough for edge devices.
• moondream2
  Small vision-language model: captions, questions, pointing and detection by text.
• Qwen2.5-VL 3B
  Strong small vision-language model: documents, charts, objects, grounding.
• CLIP ViT-B/32
  Scores how well an image matches a text, for zero-shot classification and search.

Text to speech

• Kokoro 82M
  Small, natural-sounding TTS with many voices; Junior's voice on botware uses it.
• Piper voices
  Fast offline voices for the Piper engine, one small file per language and accent.
• MeloTTS English
  English TTS with American, British, Indian and Australian accents; runs on CPU.
• Indic Parler-TTS
  Indian-language TTS (20+ languages); describe the speaker's voice in a sentence.
• MMS-TTS Hindi
  Small single-voice Hindi TTS from Massively Multilingual Speech.
• MMS-TTS Kannada
  Small single-voice Kannada TTS from Massively Multilingual Speech.
• MMS-TTS Tamil
  Small single-voice Tamil TTS from Massively Multilingual Speech.
• MMS-TTS Telugu
  Small single-voice Telugu TTS from Massively Multilingual Speech.
• MMS-TTS Malayalam
  Small single-voice Malayalam TTS from Massively Multilingual Speech.
• MMS-TTS Bengali
  Small single-voice Bengali TTS from Massively Multilingual Speech.
• MMS-TTS Marathi
  Small single-voice Marathi TTS from Massively Multilingual Speech.
• MMS-TTS English
  Small single-voice English TTS from Massively Multilingual Speech.
• SpeechT5 TTS
  English TTS where a speaker embedding picks the voice.
• Parler-TTS mini
  English TTS controlled by a text description of the voice and style.
• XTTS v2
  Multilingual voice-cloning TTS (17 languages) from a short sample.
• Bark
  Expressive multilingual speech, with laughs and pauses.

Speech recognition

• Whisper large-v3 turbo
  Fast, accurate multilingual speech-to-text and translation to English.
• Whisper tiny
  The smallest Whisper: speech-to-text on modest hardware.
• Distil-Whisper large-v3
  English-only Whisper distilled to be about 6x faster.
• Parakeet TDT 0.6B v2
  Very accurate English transcription with punctuation and timestamps.
• Moonshine tiny
  Tiny English speech recogniser built for live use on small devices.
• MMS 1B all
  Speech-to-text for 1,100+ languages with small per-language adapters.
• IndicConformer 600M
  One speech recogniser for India's 22 scheduled languages.
• Vosk
  Offline speech recognition with 50 MB models (incl. Indian English, Hindi, Telugu, Gujarati).

Audio

• openWakeWord
  Open wake-word detection ("hey robot"); train your own word from synthetic speech.
• Audio Spectrogram Transformer
  Recognises 527 everyday sounds (alarm, dog bark, glass breaking…).

Language models

• SmolLM2 360M
  Very small chat model for commands and simple tasks on-device.
• SmolLM2 1.7B
  Small chat model with good instruction following.
• Qwen3 1.7B
  Small multilingual model with a thinking mode and tool calling.
• Gemma 3 1B
  Google's smallest Gemma 3 chat model.
• Llama 3.2 1B
  Meta's small multilingual chat model for edge devices.
• Phi-4 mini
  3.8B model strong at reasoning and maths for its size.
• NLLB-200 600M
  Translates between 200 languages, including most Indian languages.
• IndicTrans2 En→Indic
  English to 22 Indian languages, built for Indian text.
• OPUS-MT English→Hindi
  Small, fast English to Hindi translator.

Robot control

• SmolVLA
  Small open VLA: fine-tune it on your own arm's demos with LeRobot (e.g. SO-101).
• π0 (LeRobot port)
  General robot policy from Physical Intelligence, ported to LeRobot.
• openpi
  Official π0 / π0-FAST models and code, with fine-tuning examples.
• OpenVLA 7B
  Open 7B VLA trained on the Open X-Embodiment robot dataset.
• GR00T N1.5
  NVIDIA's open foundation model for humanoid and arm control.
• ACT (ALOHA cube transfer)
  Action Chunking Transformer trained in simulation; a reference for training ACT on your own robot.
• Diffusion Policy (Push-T)
  Diffusion Policy trained on the Push-T task; a reference to start from.