Instructions to use L2658183739/agengt-vla with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use L2658183739/agengt-vla with LeRobot:
- Notebooks
- Google Colab
- Kaggle
Parcel Sorter v21: PI0.5 DROID + LoRA
This release contains the complete upstream lerobot/pi05_droid checkpoint at
revision 72824c0a93f00ce5bb8bedb7feb58953ba1da364 and the project v21 rank-16
LoRA adapter used for grasp-mode routing in the Parcel Sorter ROCm simulation.
本仓库包含完整的 lerobot/pi05_droid 基座快照,以及 Parcel Sorter ROCm 使用的
v21 rank-16 LoRA adapter。基座和 adapter 分目录保存,避免把上游权重与项目微调
权重混为一体。
Files / 文件结构
base/pi05-droid/
config.json
model.safetensors
policy_preprocessor.json
policy_postprocessor.json
DROID_LOCAL_BUNDLE_MANIFEST.json
adapters/v21/
adapter_model.safetensors
adapter_config.json
config.json
policy_preprocessor.json
policy_postprocessor.json
*_CONTRACT.json
train_config.json
MANIFEST_SHA256.txt
Identity / 身份与哈希
| Artifact | Identity |
|---|---|
| Base repository | lerobot/pi05_droid |
| Base revision | 72824c0a93f00ce5bb8bedb7feb58953ba1da364 |
| Base weight | 16,573,760,032 bytes |
| Base SHA-256 | 684c8b2033dcfacee9c6a83f38810ecc77df2c81aeb171bb87319da01fafcfaa |
| Adapter | LoRA rank 16, alpha 32, PEFT 0.19.1 |
| Adapter weight | 8,319,992 bytes |
| Adapter SHA-256 | beba9e950c9a3733151cd8fcce7e6bec84b1fd68fa8c57f463fdecce11ecd0a2 |
Use with Parcel Sorter ROCm / 项目加载方式
hf download L2658183739/agengt-vla --repo-type model \
--local-dir models/agengt-vla
export PARCEL_PI05_BASE_MODEL="$PWD/models/agengt-vla/base/pi05-droid"
export PARCEL_PI05_V21_CHECKPOINT="$PWD/models/agengt-vla/adapters/v21"
The project runtime uses PARCEL_PI05_BASE_MODEL to relocate the base path
stored in PEFT metadata. Do not edit adapter_config.json or create a fixed
server path.
项目运行时优先读取 PARCEL_PI05_BASE_MODEL,因此下载后无需修改 adapter metadata。
完整运行代码见
2658183739/Radeon-hackathon-2026-07。
Adapter contract / Adapter 合约
- Inputs: overhead RGB-D, left-wrist RGB-D, observable 68-D state, and task text.
- Output contract: 21-D
incremental_se3_deployment_v2, chunk size 30. - Structured heads: top suction, side suction, cooperative cradle, and primitive progress.
- Training record: 137,389 frames at 30 Hz; dataset manifest SHA-256
691182a30a7f1f628997faa7fe9b3eee16c93bb9483261eb0efc71dabd57d2ff.
Evaluation boundary / 结果边界
The recorded system is an Agent-conditioned, v21 VLA-routed,
scripted-motion hybrid. In three delivered Genesis episodes, v21 selected
side_suction by a 3/3 vote and the deterministic controller completed
0.531-0.554 m transport. Continuous VLA actions were shadowed with
applied_physics_steps=0. Strict pure-VLA parcel control remains 0/3.
当前 3 条成功轨迹只能证明 v21 参与了抓取模式路由;连续动作由脚本控制器执行, 不能写成纯 VLA 成功,也不是实机或 Sim-to-Real 结果。
Agent boundary / Agent 边界
The successful run requested gpt-5.6-luna, but the audited gateway call used
the declared gpt-5.5 fallback. The Agent supplied high-level task conditioning
only and emitted no servo, Cartesian, joint, or tool command. No API key is
stored in this model repository.
Upstream terms / 上游条款
The base files are an unmodified mirror of lerobot/pi05_droid at the revision
listed above. At verification time, that upstream model repository exposed no
model-card license field and no LICENSE file. The license: other metadata on
this repository therefore avoids assigning new terms to third-party weights.
Users must review the upstream model, PaliGemma/Gemma, LeRobot, and dataset terms
before redistribution or commercial use. Project-authored adapter metadata and
code remain subject to their respective repository notices.
Model tree for L2658183739/agengt-vla
Base model
lerobot/pi05_droid