README.md

---
library_name: lerobot
pipeline_tag: robotics
tags:
- lerobot
- dita
- libero
---

# Dita — libero_10_no_noops

This repository is a strict LeRobot conversion of `libero_long.pth` from the original
[Dita](https://github.com/RoboDita/Dita) implementation. It is configured for the
`libero_10_no_noops` action distribution and is not interchangeable with another LIBERO suite.

## Provenance and compatibility contract

- Source checkpoint: `model_ckpt/libero_long.pth`
- Source checkpoint SHA256: `4342de0cfaccc335b6288e378038db2748a703f6187217f50c340f37d49ffa7a`
- Conversion code commit: `f506ceccb9b2816ffebb024b14aeaa2b34fbd46b`
- Dataset statistics: `/home/shared/lyanam/huggingface/hub/datasets--openvla--modified_libero_rlds/snapshots/6ce6aaaaabdbe590b1eef5cd29c0d33f14a08551/libero_10_no_noops/1.0.0/dataset_statistics_3f36439171368ac499e64c677362e590ff6e0b2a87996780fc0e486ac04fd7a4.json`
- Dataset statistics SHA256: `e8b8c1dd10b983643a4aeb532b5e245437565bd12c66dbce431c59dda903c1ef`
- Image processing: one RGB primary/agentview camera, resized to 224×224, followed once by
  ImageNet mean `(0.485, 0.456, 0.406)` and std `(0.229, 0.224, 0.225)`.
- Action processing: dimensions 0–5 use this suite's `q01`/`q99`; dimension 6 is not normalized.
- Action mask: `[true, true, true, true, true, true, false]`.
- q01: `[-0.6348214149475098, -0.7741071581840515, -0.7633928656578064, -0.09749999642372131, -0.14819999992847435, -0.2742857038974762, 0.0]`
- q99: `[0.7714285850524902, 0.8464285731315613, 0.9375, 0.13928571343421936, 0.15964286029338837, 0.3246428668498993, 1.0]`
- Gripper: normalized output is thresholded with `> 0.5`, then the Cartesian dimensions are
  unnormalized, then the original Dita gripper sign flip is applied.
- Rollout: `traj_length=11`, two-frame history, `chunk_size=10`, `n_action_steps=1`,
  `num_inference_steps=10`, DDIM `prediction_type=epsilon`, and CUDA AMP by default.
- `policy.dtype` controls compute precision (`float32`, `float16`, `bfloat16`);
  parameters remain FP32. `dtype` is the sole precision setting; `use_amp` is derived internally.
- LIBERO's LeRobot environment processor performs the 180° image rotation. The policy does not
  rotate the image a second time.

## Install

```bash
pip install "lerobot_policy_dita[libero] @ https://huggingface.co/albus2024/dita_long/resolve/main/lerobot_policy_dita-2.2.0-py3-none-any.whl"
```

## Evaluate

```bash
lerobot-eval \
  --policy.path=albus2024/dita_long \
  --policy.gripper_mode=binary \
  --policy.device=cuda \
  --env.type=libero \
  --env.task=libero_10 \
  --env.max_parallel_tasks=1 \
  --eval.batch_size=1 \
  --eval.n_episodes=10 \
  --output_dir=./outputs/dita_long
```

The processor state in this repository contains the required statistics; no source `.pth` file or
local dataset metadata is needed at evaluation time.

## Validation

The converter checks all parameter names and shapes, loads with `strict=True`, and checks the
saved artifact files. Run the project's parity and converted-checkpoint verification scripts
separately before using the policy; conversion alone does not establish rollout parity.

## Runtime 2.1.0 training configuration

Install the **2.2.0 wheel above** for this revision. The new training fields
require this runtime. The policy weights, processor files and action statistics
are preserved; this is a configuration/runtime update, not a newly trained model.

| Setting | Value |
| --- | --- |
| `freeze_backbone` | `false` |
| `gradient_checkpointing` | `true`, only during gradient-enabled training |
| `optimizer_lr` | `0.0005` |
| `optimizer_backbone_lr_scale` | `0.1` |
| AdamW betas / epsilon | `[0.9, 0.95]` / `1e-8` |
| Weight decay | `0.05`; bias and one-dimensional parameters use `0` |
| Gradient clipping | Disabled (`optimizer_grad_clip_norm=0`) |
| LR schedule | 1000-step warmup, cosine to 100000 steps, minimum scale `0.01` |
| `lora_dropout` | `0.0` |

Inference automatically skips ViT and QFormer activation-checkpoint wrappers.
`policy.dtype` remains `float16` by default; BF16 and FP32 are also supported.
CLIP text encoding remains frozen. The gripper mode remains
`binary`; see [gripper modes](GRIPPER_MODES.md).

### Fine-tuning and LoRA

Use the following for full fine-tuning (choose an appropriate target dataset):

```bash
lerobot-train \
  --policy.path=albus2024/dita_long \
  --policy.device=cuda \
  --policy.dtype=bfloat16 \
  --policy.input_features=null \
  --policy.push_to_hub=false \
  --dataset.repo_id=YOUR_USERNAME/YOUR_DATASET \
  --output_dir=outputs/train/dita_long_full \
  --steps=100000
```

For LoRA, install the optional dependency and add the adapter arguments:

```bash
pip install "lerobot_policy_dita[lora] @ https://huggingface.co/albus2024/dita_long/resolve/main/lerobot_policy_dita-2.2.0-py3-none-any.whl"
# Add to lerobot-train and choose a fresh output directory:
# --peft.method_type=LORA --peft.r=32
```

LoRA targets every Dita linear layer, including the ViT, QFormer/FiLM and action
head. Only adapter weights are trained. The default alpha is twice the rank;
initialization is Gaussian for A and zero for B. Use LeRobot's native policy
factory/evaluation entry points to load saved adapters together with their base
model. Do not set `policy.use_peft=true` when starting from a base checkpoint.

The LR schedule keeps its configured 100000-step period even in shorter runs;
the first update has zero LR. Set `policy.scheduler_decay_steps` explicitly for
a different period. These defaults follow the upstream LIBERO recipe, including
the higher `5e-4` peak LR for LIBERO-Long; DROID uses `1e-4` as a fine-tuning
starting point and can be overridden for the target task. When adapting DROID
to LIBERO, pass `--policy.gripper_mode=binary`.

See [runtime validation](runtime_validation_report.json) for the checks performed
with this release and [source hashes](runtime_source_manifest.json) for the exact
packaged code. The previous [runtime 2.0.1 report](history/runtime-2.0.1/runtime_validation_report.json)
is archived. Short training/inference checks do not establish convergence or
robot task success.


## Runtime 2.2.0 image preprocessing

Install the **2.2.0 wheel linked above** before loading this revision. The saved
config now explicitly contains `image_preprocessing_device=cpu`. The weights,
processor files and action statistics remain the same.

Pixel validation always uses one min/max reduction on the model device: on CUDA
it runs on the GPU even when image resizing runs on the CPU. NaN, infinities and
pixels outside `[0, 1]` are rejected. There is no validation-device option.

`policy.image_preprocessing_device` controls FP32 antialiased bilinear resizing
and ImageNet normalization in both training and inference:

| Value | Behavior |
| --- | --- |
| `cpu` | CPU resize/normalization, preserving the reference numerical path |
| `policy` | Resize/normalization on the actual model device, including local CUDA ranks |

To fine-tune using GPU preprocessing, add this CLI override:

```bash
--policy.image_preprocessing_device=policy
```

This overrides an initialization saved with `cpu`. The fine-tuned full-policy
or LoRA checkpoint saves `policy`, which is restored for inference. Use the same
saved setting for training and deployment. `policy.dtype` controls model AMP;
image preprocessing remains FP32 for all supported compute dtypes.

With fixed RNG and deterministic kernels, the new `cpu` mode exactly matched the
previous implementation's loss, all parameter gradients, and inference actions
on the checked real-data batch in FP32, BF16 and FP16. Switching existing weights
to GPU resizing produces measurable differences, especially under AMP; no
rollout-equivalence or convergence claim is made. This published configuration
therefore remains `cpu`.

See [per-model numerical validation](runtime_validation_report.json),
[training and deployment instructions](TRAINING.md), and the
[archived runtime 2.1.0 configuration](history/runtime-2.1.0/config.json).