Skip to content
Robotics

khanhnd61/pop-pretrain-so101-community

PoP policy checkpoints for the SO-101 robot

2 min read

This repository provides pretraining checkpoints and fine-tuned policy weights for the PoP (Pointer Policy) robotic control system designed for the SO-101 manipulator. Instead of text instructions, the model takes a 2D object coordinate, a 2D target coordinate, and a short skill code, predicting joint actions based on images from fixed and wrist-mounted cameras. Built on a DINOv2 ViT-S/14 visual encoder and an adaLN-zero transformer decoder (roughly 28M parameters) by khanhnd61, this project suits robotics engineers and researchers exploring manipulator control without large language models.

Key features

  • Pretrained on 10,639 episodes: The visual encoder and control head (27.74M parameters) were trained via behavioral cloning on the HuggingFaceVLA/community_dataset_v2 dataset.
  • Pointers instead of text: The model accepts 2D object and target points with Fourier features and a skill code instead of text task descriptions.
  • Baseline comparison: The card includes L1 validation error benchmarks for PoP alongside a text-conditioned baseline using a T5-Efficient-Tiny encoder on the identical dataset.
  • Four manipulation tasks: Towel folding, drawer opening with block placement, cup stacking, and mug pushing toward a target.

What you can do

  • Test on real SO-101 hardware: Run the fine-tuned checkpoints on an SO-101 manipulator to evaluate real-world policy performance, as the card currently provides demonstration validation errors only.
  • Compare conditioning methods: Use the published weights and logs to analyze trajectory prediction accuracy when specifying goals via coordinate points versus text prompts.
  • Build custom tasks: Use the pretrained step_200000.pt checkpoint as a foundation to fine-tune on your own 6-axis manipulator demonstrations.

Example

import torch
ckpt = torch.load("finetune/261121-so101-pop-s0/step_040000.pt", map_location="cpu", weights_only=False)
ckpt["ema"]           # веса модели (state dict)
ckpt["norm"]          # нормализация состояний и действий
ckpt["instructions"]  # токенизированные инструкции (для текстовых прогонов)
ckpt["config"]        # параметры обучения
Read moreDetails
The best L1 validation error at step 40k for PoP is 0.292 (0.284 without pretraining), with minimum error reached between steps 4k and 6k (0.271–0.278).
Base architecture: DINOv2 ViT-S/14 (22.06M) plus an adaLN-zero transformer decoder (5.68M), trained in bf16 with a batch size of 64.
Limitations: Model architecture code is not included in the repository, checkpoints have not yet been validated via real-world robot deployment, and overfitting occurs by step 40k.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # khanhnd61/pop-pretrain-so100-community
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download khanhnd61/pop-pretrain-so100-community --local-dir ~/models/pop-pretrain-so100-community
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of dinov2-small (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

khanhnd61/pop-pretrain-so100-community
Author
khanhnd61
Category
Robotics
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments