Scripts and Benchmark Results for Simpler Environments?

[zbzhu99]

Thanks for your great work on EMMA-X and for open-sourcing it! I noticed that the scripts and results are provided on real robot testing, but I saw references to simpler env in the codebase (experiments/SimplerEnv-OpenVLA). Are there any evaluation scripts or benchmark results for testing EMMA-X in Simpler environments?

[yxy-2002]

i have the same question. I also want to deploy this model in Simpler_env simulator.

[CostaliyA]

the same @Emrys-Hong

[Akila-Ayanthi]

Any updates?

[CostaliyA]

I implemented my own evaluation of emma-x on Simplerenv, but the results were not ideal.

[CostaliyA]

`from typing import Optional, Sequence
import os
import matplotlib.pyplot as plt
import numpy as np
from transforms3d.euler import euler2axangle
from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image
import torch
import cv2 as cv
import sys

class EMMAXInference:
def init(
self,
saved_model_path: str = "openvla/openvla-7b",
unnorm_key: Optional[str] = None,
policy_setup: str = "widowx_bridge",
horizon: int = 1,
pred_action_horizon: int = 1,
exec_horizon: int = 1,
image_size: list[int] = [224, 224],
action_scale: float = 1.0,
) -> None:
os.environ["TOKENIZERS_PARALLELISM"] = "false"
if policy_setup == "widowx_bridge":
unnorm_key = "bridge_orig" if unnorm_key is None else unnorm_key
self.sticky_gripper_num_repeat = 1
elif policy_setup == "google_robot":
unnorm_key = "bridge_orig" if unnorm_key is None else unnorm_key
self.sticky_gripper_num_repeat = 15
else:
raise NotImplementedError(
f"Policy setup {policy_setup} not supported for octo models. The other datasets can be found in the huggingface config.json file."
)
self.policy_setup = policy_setup
self.unnorm_key = unnorm_key

    print(f"*** policy_setup: {policy_setup}, unnorm_key: {unnorm_key} ***")
    self.processor = AutoProcessor.from_pretrained(saved_model_path, trust_remote_code=True)
    self.vla = AutoModelForVision2Seq.from_pretrained(
    saved_model_path,
    attn_implementation="flash_attention_2",  # [Optional] Requires `flash_attn`
    torch_dtype=torch.bfloat16, 
    low_cpu_mem_usage=True, 
    trust_remote_code=True
    ).to("cuda")

    self.image_size = image_size
    self.action_scale = action_scale
    self.horizon = horizon
    self.pred_action_horizon = pred_action_horizon
    self.exec_horizon = exec_horizon

    self.sticky_action_is_on = False
    self.gripper_action_repeat = 0
    self.sticky_gripper_action = 0.0
    self.previous_gripper_action = None

    self.task = None
    self.task_description = None
    self.num_image_history = 0
    self.reasoning_content_list = []

def reset(self, task_description: str) -> None:
    outputs = []
    for i, reasoning_content in enumerate(self.reasoning_content_list, start=1):
        outputs.append(f"<think>{i}.{reasoning_content}</think>\n")
    sys.stdout.writelines(outputs)
    sys.stdout.flush()
    self.reasoning_content_list=[]
    self.task_description = task_description
    self.num_image_history = 0

    self.sticky_action_is_on = False
    self.gripper_action_repeat = 0
    self.sticky_gripper_action = 0.0
    self.previous_gripper_action = None

def step(
    self, image: np.ndarray, task_description: Optional[str] = None, *args, **kwargs
) -> tuple[dict[str, np.ndarray], dict[str, np.ndarray]]:
    """
    Input:
        image: np.ndarray of shape (H, W, 3), uint8
        task_description: Optional[str], task description; if different from previous task description, policy state is reset
    Output:
        raw_action: dict; raw policy action output
        action: dict; processed action to be sent to the maniskill2 environment, with the following keys:
            - 'world_vector': np.ndarray of shape (3,), xyz translation of robot end-effector
            - 'rot_axangle': np.ndarray of shape (3,), axis-angle representation of end-effector rotation
            - 'gripper': np.ndarray of shape (1,), gripper action
            - 'terminate_episode': np.ndarray of shape (1,), 1 if episode should be terminated, 0 otherwise
    """
    if task_description is not None:
        if task_description != self.task_description:
            self.reset(task_description)

    assert image.dtype == np.uint8
    image = self._resize_image(image)

    image: Image.Image = Image.fromarray(image)
    prompt = task_description


    # predict action (7-dof; un-normalize for bridgev2)
    prompt, image = self.processor.get_prompt(prompt, image)

    inputs = self.processor(prompt, image).to("cuda", dtype=torch.bfloat16)
    raw_actions,reasoning = self.vla.generate_actions(inputs, self.processor.tokenizer, do_sample=False, max_new_tokens=1024)

    raw_actions = raw_actions[None, ...] 
    self.reasoning_content_list.append(reasoning)
    # print(f"*** raw actions {raw_actions} ***")

    raw_action = {
        "world_vector": np.array(raw_actions[0, :3]),
        "rotation_delta": np.array(raw_actions[0, 3:6]),
        "open_gripper": np.array(raw_actions[0, 6:7]),  # range [0, 1]; 1 = open; 0 = close
    }

    # process raw_action to obtain the action to be sent to the maniskill2 environment
    action = {}
    action["world_vector"] = raw_action["world_vector"] * self.action_scale
    action_rotation_delta = np.asarray(raw_action["rotation_delta"], dtype=np.float64)
    roll, pitch, yaw = action_rotation_delta
    action_rotation_ax, action_rotation_angle = euler2axangle(roll, pitch, yaw)
    action_rotation_axangle = action_rotation_ax * action_rotation_angle
    action["rot_axangle"] = action_rotation_axangle * self.action_scale

    if self.policy_setup == "google_robot":
        current_gripper_action = raw_action["open_gripper"]
        if self.previous_gripper_action is None:
            relative_gripper_action = np.array([0])
        else:
            relative_gripper_action = self.previous_gripper_action - current_gripper_action
        self.previous_gripper_action = current_gripper_action

        if np.abs(relative_gripper_action) > 0.5 and (not self.sticky_action_is_on):
            self.sticky_action_is_on = True
            self.sticky_gripper_action = relative_gripper_action

        if self.sticky_action_is_on:
            self.gripper_action_repeat += 1
            relative_gripper_action = self.sticky_gripper_action

        if self.gripper_action_repeat == self.sticky_gripper_num_repeat:
            self.sticky_action_is_on = False
            self.gripper_action_repeat = 0
            self.sticky_gripper_action = 0.0

        action["gripper"] = relative_gripper_action

    elif self.policy_setup == "widowx_bridge":
        action["gripper"] = 2.0 * (raw_action["open_gripper"] > 0.5) - 1.0

    action["terminate_episode"] = np.array([0.0])

    return raw_action, action

def _resize_image(self, image: np.ndarray) -> np.ndarray:
    image = cv.resize(image, tuple(self.image_size), interpolation=cv.INTER_AREA)
    return image

def visualize_epoch(
    self, predicted_raw_actions: Sequence[np.ndarray], images: Sequence[np.ndarray], save_path: str
) -> None:
    images = [self._resize_image(image) for image in images]
    ACTION_DIM_LABELS = ["x", "y", "z", "roll", "pitch", "yaw", "grasp"]

    img_strip = np.concatenate(np.array(images[::3]), axis=1)

    # set up plt figure
    figure_layout = [["image"] * len(ACTION_DIM_LABELS), ACTION_DIM_LABELS]
    plt.rcParams.update({"font.size": 12})
    fig, axs = plt.subplot_mosaic(figure_layout)
    fig.set_size_inches([45, 10])

    # plot actions
    pred_actions = np.array(
        [
            np.concatenate([a["world_vector"], a["rotation_delta"], a["open_gripper"]], axis=-1)
            for a in predicted_raw_actions
        ]
    )
    for action_dim, action_label in enumerate(ACTION_DIM_LABELS):
        # actions have batch, horizon, dim, in this example we just take the first action for simplicity
        axs[action_label].plot(pred_actions[:, action_dim], label="predicted action")
        axs[action_label].set_title(action_label)
        axs[action_label].set_xlabel("Time in one episode")

    axs["image"].imshow(img_strip)
    axs["image"].set_xlabel("Time in one episode (subsampled)")
    plt.legend()
    plt.savefig(save_path)

`You can refer to this.

[Akila-Ayanthi]

Thanks. I will check this.
How did your results look?

[CostaliyA]

Thanks. I will check this. How did your results look?

Not GOOD T.T. You can refer to https://arxiv.org/pdf/2508.07917 table 1. This result aligns with mine.