Fine-tuned GR00T-N1.5 victim policy trained on sequentially perturbed LIBERO-10 demonstrations.
Open-weight model · Robotics
groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k
by Zhoumiaosen zhoumiaosen/groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k
groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k is an open-weight model for robotics from Zhoumiaosen, released under Apache License 2.0. It has 2.4B parameters. At 16-bit it needs about 5.8 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
GR00T-N1.5-3B fine-tuned with LeRobot (0.4.1) on the 379 LIBERO-10 (LIBERO-Long) demonstrations, on a single RTX 4090. This is the final 60,000-step checkpoint. It is a color-jitter ablation of an RGB-vs-grayscale study.
Parameters2.4B
Context—
Weights7.0 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k (2.4B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 4.8 GB | 5.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 2.4 GB | 2.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 1.2 GB | 1.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
Model Card
By Zhoumiaosen, published under apache-2.0, revision 6ceef7a72157.
GR00T-N1.5-3B fine-tuned with LeRobot (0.4.1) on the 379 LIBERO-10 (LIBERO-Long) demonstrations, on a single RTX 4090. This is the final 60,000-step checkpoint. It is a color-jitter ablation of an RGB-vs-grayscale study. The recipe is identical to except that training-time color jitter is turned off. It reaches 66% on LIBERO-10 under default lighting. It is much less robust to lighting changes than the jittered RGB and grayscale models. zhoumiaosen/libero10rgb: the 379 LIBERO-10 episodes of openvla/modifiedliberorlds, converted to LeRobot v3.0 at 10 fps. That is 101,469 frames. observation.state (8-D: eef position, axis-angle, 2 gripper joints). Output: 7-D action (eef delta + gripper).…
Read Zhoumiaosen's full model card
GR00T-N1.5-3B fine-tuned on LIBERO-10, RGB without color jitter (batch 16, 60k steps)
GR00T-N1.5-3B fine-tuned with LeRobot (0.4.1) on the 379 LIBERO-10 (LIBERO-Long) demonstrations, on a single RTX 4090. This is the final 60,000-step checkpoint.
It is a color-jitter ablation of an RGB-vs-grayscale study. The recipe is identical to
zhoumiaosen/groot-n1p5-lora-libero10-rgb-bs16-60k
except that training-time color jitter is turned off. It reaches 66% on LIBERO-10 under default lighting. It is
much less robust to lighting changes than the jittered RGB and grayscale models.
Training data
zhoumiaosen/libero_10_rgb: the 379 LIBERO-10 episodes of
openvla/modified_libero_rlds, converted to LeRobot v3.0 at
10 fps. That is 101,469 frames.
- Inputs:
observation.images.image(main camera, 256×256 RGB),observation.images.image2(wrist camera),observation.state(8-D: eef position, axis-angle, 2 gripper joints). Output: 7-Daction(eef delta + gripper).
Training recipe
lerobot-train --policy.type=groot --policy.lora_rank=8 --policy.tune_projector=false \
--dataset.repo_id=zhoumiaosen/libero_10_rgb --dataset.image_transforms.enable=false \
--batch_size=16 --steps=60000 --seed=1000
- Batch 16 × 60k steps is about 960k samples, roughly 9.5 epochs. Learning rate 1e-4 with the default warmup and cosine decay.
- No image augmentation. The jittered models used ColorJitter with brightness 0.7–1.3, contrast 0.6–1.4, saturation 0.5–1.5 and hue ±0.08.
- The projector is frozen.
lora_rank=8does not reduce the trainable parameters in LeRobot's GR00T integration: the diffusion action head (~750M parameters) is fully fine-tuned. - Final training loss was about 0.03.
Evaluation (LIBERO-10)
200 episodes per row: 10 tasks × 20 distinct scenes, each set by a fixed per-episode seed. The scenes are the same for
every model compared here. They are LIBERO's seeded random object placements, not LIBERO's stored init states:
LeRobot 0.4.1's LiberoEnv applies the init state before LIBERO's reset rebuilds the scene, so the stored state is
discarded. Every action chunk is executed in full, and episodes are capped at 520 steps.
Lighting changes are applied at test time to both scene lights and the renderer's headlight: - dim: all intensities ×0.4 - bright: ×1.8 - warm: orange per-channel gain (1.0, 0.65, 0.35) - cool: blue per-channel gain (0.45, 0.7, 1.0)
| Lighting (60k steps) | RGB, no jitter (this model) | RGB + jitter | Gray + jitter |
|---|---|---|---|
| default | 66.0% | 71.0% | 74.0% |
| dim ×0.4 | 51.5% | 71.0% | 75.0% |
| bright ×1.8 | 62.5% | 67.0% | 70.5% |
| warm | 66.0% | 70.5% | 77.0% |
| cool | 59.5% | 73.0% | 74.0% |
- Default lighting: across checkpoints this model scores 63.5% at 20k, 71.0% at 40k and 66.0% at 60k. None of these differ significantly from the jittered models (exact McNemar, p ≥ 0.064).
- Dim and blue light: it falls significantly below both jittered models (p = 3.8e-6 and 7.3e-4 vs RGB + jitter; 2.3e-8 and 0.0015 vs gray). These survive a Bonferroni correction over 14 tests.
- Its own drop in dim light: 66.0% to 51.5%, p = 4.2e-4.
- Conclusion: color jitter, not the pretrained encoder, is what makes the RGB policy robust to lighting.
Per-task successes (out of 20), default lighting: task 0: 0, 1: 15, 2: 19, 3: 15, 4: 8, 5: 18, 6: 15, 7: 13, 8: 12, 9: 17. Each model is a single training run (seed 1000).
Identity and Version
- Repository
- zhoumiaosen/groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k
- Publisher
- Zhoumiaosen
- Task
- Robotics
- Modality
- Control
- Library
- lerobot
- Parameters
- 2.4B parameters
- Languages
- Not stated by the source
- Revision
- 6ceef7a72157d4a1bd88f3b4a17e6f1a55de0811
- First published
- 2026-10-18
- Last updated
- 2026-10-18
Files and Weights
9 files, 7.0 GB in total. The weights are 3 files totalling 7.0 GB in safetensors.
Weights3 files · 7.0 GB
Configuration4 files · 8.8 KB
Documentation1 file · 3.9 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 7.0 GB | 4998bfaa23e6 |
| policy_postprocessor_step_0_groot_action_unpack_unnormalize_v1.safetensors | Weights | 8.6 KB | 4630164087e1 |
| policy_preprocessor_step_2_groot_pack_inputs_v3.safetensors | Weights | 8.6 KB | 4630164087e1 |
| config.json | Configuration | 2.2 KB | — |
| policy_postprocessor.json | Configuration | 449 B | — |
| policy_preprocessor.json | Configuration | 1.2 KB | — |
| train_config.json | Configuration | 5.0 KB | — |
| README.md | Documentation | 3.9 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 7.0 GB
Download from Zhoumiaosen
Released by Zhoumiaosen through its official repository on Hugging Face. Read the license.
Built From
- Derived from nvidia/GR00T-N1.5-3B
- Trained on (disclosed) zhoumiaosen/libero_10_rgb
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 7.0 GB |
| 16-bit | 4.8 GB |
| 8-bit | 2.4 GB |
| 4-bit | 1.2 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k
How much GPU memory does groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k need?
About 5.8 GB at 16-bit and 1.4 GB at 4-bit: the weights (2.4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k commercially?
Yes. groot-n1p5-lora-libero10-rgb-nojitter-bs16-60k is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
Similar Models
압축 가능성 라벨로 게이트를 학습한 액션 양자화(ATQ) 체크포인트다. 하나의 정책이 미세(1x) 디코더와 압축 디코더를 함께 갖고, VLM 라벨에서 배운 conf 가 둘 중 어느 군을 쓸지 고른다. 라우터는 그 군 안에서 horizon 만 고른다. (VLM 전용, 접촉 열 없음 · 16,286행 · 1,693에피 · stride 16) moeexperthorizons = [16, 9, 5, 8] · confthreshold(tau) = 0.55 · discreteactiondims = [6] (그리퍼는 절대 명령이라 · actionmergereduction = sum 회전 병합은 SO(3) 다(rotationmergespec 이 config 에 있다). 압축 블록의 회전 다시 정규화한다. scipy 대조 각도 오차 1e-14도. LIBERO 는 5 fine 스텝마다 재계획한다. 압축 행 하나는 fine 액션 2~3개의 합이므로 같은 배속이 되고, 배속을 움직이는 손잡이는 conf 게이트 하나다: 넘으므로 OSC 팔 컨트롤러의 입력 클립을 제거한 조건에서 평가했다(그리퍼 그대로). 게이트는 벤치마크가 실제로 깨지는 순서를 따른다 -- 압축에 강한 liberoobject 를 가장 많이 압축하고, 2배에서 -0.160 으로 무너지는 liberospatial 은 거의 압축하지 않는다. 브랜치 jimin-dev-label-gated.…
압축 가능성 라벨로 게이트를 학습한 액션 양자화(ATQ) 체크포인트다. 하나의 정책이 미세(1x) 디코더와 압축 디코더를 함께 갖고, VLM 라벨에서 배운 conf 가 둘 중 어느 군을 쓸지 고른다. 라우터는 그 군 안에서 horizon 만 고른다. (VLM 전용, 접촉 열 없음 · 16,286행 · 1,693에피 · stride 16) moeexperthorizons = [16, 7, 3, 8] · confthreshold(tau) = 0.55 · discreteactiondims = [6] (그리퍼는 절대 명령이라 · actionmergereduction = sum 회전 병합은 SO(3) 다(rotationmergespec 이 config 에 있다). 압축 블록의 회전 다시 정규화한다. scipy 대조 각도 오차 1e-14도. LIBERO 는 5 fine 스텝마다 재계획한다. 압축 행 하나는 fine 액션 2~3개의 합이므로 같은 배속이 되고, 배속을 움직이는 손잡이는 conf 게이트 하나다: 넘으므로 OSC 팔 컨트롤러의 입력 클립을 제거한 조건에서 평가했다(그리퍼 그대로). 게이트는 벤치마크가 실제로 깨지는 순서를 따른다 -- 압축에 강한 liberoobject 를 가장 많이 압축하고, 2배에서 -0.160 으로 무너지는 liberospatial 은 거의 압축하지 않는다. 브랜치 jimin-dev-label-gated.…
NVIDIA Isaac GR00T N1.7 is an open foundation model for generalized humanoid robot reasoning and skills. This cross-embodiment model takes multimodal input, including language and images, to perform manipulation tasks in diverse environments. Developers and researchers can post-train GR00T N1.7 with real or synthetic data for their specific humanoid robot or task. Isaac GR00T N1.7 is the medium-sized version of our model built using pre-trained vision and language encoders, and uses a flow matching action transformer to model a chunk of actions conditioned on vision, language and proprioception. A detailed description of the Isaac GR00T N1.X architecture is provided in the GROOT N1 White…
NVIDIA Isaac GR00T N1.7 is an open foundation model for generalized humanoid robot reasoning and skills. This cross-embodiment model takes multimodal input, including language and images, to perform manipulation tasks in diverse environments. Developers and researchers can post-train GR00T N1.7 with real or synthetic data for their specific humanoid robot or task. Isaac GR00T N1.7 is the medium-sized version of our model built using pre-trained vision and language encoders, and uses a flow matching action transformer to model a chunk of actions conditioned on vision, language and proprioception. A detailed description of the Isaac GR00T N1.X architecture is provided in the GROOT N1 White…
Independent further-training of the new sim-only 20,000-step policy on 10 real10 episodes for 5,000 additional steps. Both real10 and Cosmos10 branches initialize from the same sim-only parent; neither continues from the other branch. - Native parent processors and statistics retained, with stock clipping; no unclipping ablation. - Visual tuning enabled, language tuning disabled, BF16 autocast with FP32 parameters; image augmentation enabled. - Checkpoints at 2,500 and 5,000 additional steps. This repository contains the final 5,000-step continuation. The native policy and processor files must remain together. See training-provenance.json for parent weight hash and exact provenance.…