yqi19/xvla-inference

X-VLA Inference & ManiSkill Conflict Evaluation

This repository contains the inference code and ManiSkill conflict evaluation environment for X-VLA (Soft-Prompted Transformer as a Scalable Cross-Embodiment Vision-Language-Action Model), evaluated on OOD conflict experiments.

Note: XVLA achieved 0% task success on all 10 conflict experiments (color_object evaluated; gripper retracted without interaction ~90% of the time). This repo is provided for reproducibility and as a reference implementation. See the companion repos for models that achieved higher success: genie-inference-maniskill.

---

Repository Structure

xvla-inference/
├── xvla/                              # X-VLA inference code
│   ├── main.py                        # Main rollout script (conflict eval)
│   ├── deploy.py                      # X-VLA FastAPI inference server launcher
│   ├── run_ood_experiment_inference.sh # Batch OOD sweep script
│   ├── vlm_eval_xvla_conflicts.py     # VLM eval (Gemini-2.5-Flash) of rollout videos
│   ├── models/                        # Model architecture (Florence2 + X-VLA)
│   ├── deploy/                        # X-VLA-Pt base pretrained model config
│   ├── conflict_stats.json            # Action/state normalization stats
│   ├── requirements.txt               # Python dependencies
│   └── finetune_readme.md             # Fine-tuning guide
│
└── maniskill_conflict/                # ManiSkill conflict environment
    ├── mani_skill/                    # Modified ManiSkill package (VerbObjectColor-v1)
    ├── conflict_experiment/           # Pair generation & experiment utilities
    └── collection_strategy/           # Training vocabulary & task language

---

Action Space

X-VLA outputs action shape (30, 20) — 30 steps × 20 dims — but only the first 8 dims are meaningful joint-space targets:

Component	Dims	Description
Joint targets	7	q0..q6 absolute joint positions
Gripper	1	Gripper open/close target
Padding	12	Zero-padded (training artifact)
Total	20	per action step

• Control mode: pd_joint_pos (absolute joint position)
• The server outputs z-score normalized actions; main.py unnormalizes using conflict_stats.json
• State input: 8-dim joint state, z-score normalized, zero-padded to 20 dims before sending to server

---

Conflict Experiments

Ten pairwise factor conflict experiments were evaluated (each with 400 OOD episodes):

Experiment	Factor 1	Factor 2
color_object	color	shape
color_size	color	size
color_spatial	color	spatial
size_object	size	shape
spatial_object	spatial	shape
spatial_size	spatial	size
verb_color	verb	color
verb_object	verb	shape
verb_size	verb	size
verb_spatial	verb	spatial

XVLA result: 0% success across all experiments. The policy (fine-tuned from X-VLA-Pt checkpoint at step 30000) did not develop directional grasping behavior in the conflict environment — the gripper retracted without interacting with objects ~90% of the time.

---

Setup

1. Install dependencies

git clone https://huggingface.co/datasets/yqi19/xvla-inference
cd xvla-inference

# Install ManiSkill conflict environment
pip install -e maniskill_conflict/

# Install X-VLA dependencies
pip install -r xvla/requirements.txt

2. Download checkpoint

Download the X-VLA fine-tuned checkpoint from yqi19/xvla:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="yqi19/xvla", local_dir="./xvla_checkpoint")

Place the checkpoint at xvla/checkpoints/color_object/ckpt-30000/.

3. Launch inference server

cd xvla
python deploy.py --model_path checkpoints/color_object/ckpt-30000 --port 8010

4. Run conflict evaluation

# Single pair
python xvla/main.py \
    --host localhost --port 8010 \
    --experiment color_object --pair-i 0 --pair-j 1 --run-type color \
    --num-episodes 20

# Full sweep (all pairs for one experiment)
bash xvla/run_ood_experiment_inference.sh color_object 8010

---

VLM Evaluation

After collecting rollout videos, evaluate factor dominance using Gemini-2.5-Flash:

export NVIDIA_API_KEY=<your_api_key>
export XVLA_EXPERIMENTS_DIR=xvla/data/conflict_xvla/experiments
export XVLA_RESULTS_DIR=results
export XVLA_VLM_OUT_DIR=vlm_eval_output
python xvla/vlm_eval_xvla_conflicts.py color_object

Note: Since XVLA had 0% task success, there are no meaningful videos to evaluate — the VLM script is included for completeness.

---

Citation

@article{zheng2025x,
  title   = {X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model},
  author  = {Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and others},
  journal = {arXiv preprint arXiv:2510.10274},
  year    = {2025}
}