SIMPLER¶

Eden ports the four WidowX (BridgeData V2) tasks of SIMPLER, the real-to-sim evaluation suite for generalist manipulation policies. Read this page to evaluate a VLA such as Octo, OpenPI or GR00T on them, or to generate demonstrations. The code is in eden/tasks/simpler/.

The WidowX 250 in simpler_put_carrot_on_plate

The tasks¶

Task (`--task`)	SIMPLER task	Instruction	Success metric
`simpler_put_spoon_on_table_cloth`	`PutSpoonOnTableClothInScene`	"put spoon on table cloth"	`placement`: the spoon on the cloth (`object_a_is_on_b`)
`simpler_put_carrot_on_plate`	`PutCarrotOnPlateInScene`	"put carrot on plate"	`placement`: the carrot on the plate
`simpler_stack_green_cube_on_yellow_cube`	`StackGreenCubeOnYellowCubeInScene`	"put green cube on yellow cube"	`placement`: the green cube on the yellow one
`simpler_put_eggplant_in_basket`	`PutEggplantInBasketScene`	"put eggplant in basket"	`placement`: the eggplant inside the sink's basket box (`object_in_region`)

The tasks are built for evaluation and demonstration, not RL: they define a success metric but no rewards, and they default to one env and SIMPLER's 5 Hz control. Every scene's objects are source (what is moved) and target (where it goes), so a metric or a demonstrator reads the same names on every task.

What matches SIMPLER¶

The robot. A system-identified WidowX 250 (eden.tasks.simpler.base_config.DEFAULT_ROBOT): its kp/kd tables come from fitting the real BridgeData V2 arm, not the robot's stock gains, and it is gravity-compensated.
The frame. The world frame is the arm's base_link, as in SIMPLER, which mounts its cameras on that link.
The camera. SIMPLER's base-mounted 3rd_view_camera pose for each scene, with the Logitech C920 intrinsics the BridgeData V2 demonstrations were recorded with (C920_INTRINSIC).
The real-image backdrop. Each camera composites the scene over SIMPLER's real "inpainting" photo (real_inpainting/bridge_real_eval_1.png for the table tasks, real_inpainting/bridge_sink.png for the sink), through a BackgroundImageComposite camera overlay. The ground plane, and on the sink task the sink, are drawn by the photo rather than rendered.
The layouts. The three table tasks place their objects from SIMPLER's own episode grids (bridge_table.config.simpler_layouts, applied by the PlaceFromLayouts event): every ordered pair of corners of a square round the grid centre, once per source yaw, in SIMPLER's episode order.
Control. 5 Hz (sim_dt=0.005, decimation=40), a DifferentialIKController on gripper_link (ee_controller) and a binary gripper (gripper_controller).

Where Eden differs:

Torsional friction is on, with a 2 cm contact patch on the fingers (default_geoms_friction_torsional). Without it a spoon pinched at one point spins about the line between the fingertips and hangs below the grasp.
The sink is fitted by eye. Its pose and scale (SINK_POS, SINK_SCALE) were aligned to the backdrop photo through the task camera, and the eggplant spawns on the basin floor's +y half so a rolled one stays reachable.
The sim step is 5 ms (200 Hz), where SIMPLER used 2 ms.

Evaluating a generalist policy¶

Three remote-policy clients have a setup for SIMPLER's WidowX:

Client	Setup	Notes
`OctoPolicy` (`eden.extensions.policies.octo.policy`)	(single setup)	one action per request
`PiPolicy` (`eden.extensions.policies.pi.policy`)	`policy_setup="widowx_bridge"`	Bridge EE-delta
`Gr00tPolicy` (`eden.extensions.policies.gr00t.policy`)	`policy_setup="simpler_env_widowx"`	256 × 256 images

Each client brings its own observation groups and action terms (its gripper is a GripperDecisionController reading the server's open/close convention, not the task's BinaryJointController), so replace both on the config before building. The clients need the policy-client extra (pip install 'eden-world[policy-client]', or uv sync --inexact --extra policy-client in a checkout).

import eden as en
import genesis as gs
from eden.envs.base import env_class_for_config
from eden.extensions.policies.gr00t.policy import Gr00tPolicy
from eden.options.managers.actions import ActionManagerOptions
from eden.options.managers.observations import ObservationManagerOptions
from eden.tasks import TASK_REGISTRY

en.init(backend=gs.gpu)
cfg = TASK_REGISTRY.build("simpler_put_carrot_on_plate")
cfg.observation_options = ObservationManagerOptions(**Gr00tPolicy.observation_options(policy_setup="simpler_env_widowx"))
cfg.action_options = ActionManagerOptions(**Gr00tPolicy.action_options(policy_setup="simpler_env_widowx"))
env = env_class_for_config(cfg).from_config(cfg, eval_mode=True)
env.build()

policy = Gr00tPolicy(env, Gr00tPolicy.configure(host="localhost", port=5555, policy_setup="simpler_env_widowx"))
policy.set_task(cfg.language_instructions[0])

for episode in range(5):
    obs, _ = env.reset()
    policy.reset_server()
    for _ in range(100):
        obs, *_ = env.step(policy.step(obs))

eden inference --task simpler_put_carrot_on_plate --policy gr00t runs the same loop from the command line with the client's defaults. The PiPolicy and Gr00tPolicy setups check the env when the client is constructed and refuse one that does not step at SIMPLER's 5 Hz, or whose observation and action terms do not match the setup. See Remote policies.

Scripted demonstrations¶

Every SIMPLER task has an eden datagen demonstrator (eden/tasks/simpler/scripted_demos.py): a keyframe tower controller for the cube stack and a put-on tool routine for the other three. The put-on demonstrators can also run camera-only (--camera_only), locating the objects through the task camera.

eden datagen --task simpler_stack_green_cube_on_yellow_cube --num_envs 16 --num_episodes 64
eden datagen --task simpler_put_carrot_on_plate --camera_only

Success rates and what each demonstrator does are in Scripted demonstrations.

Assets¶

The objects (eden.options.objects.simpler: SimplerCarrot, SimplerSpoon, SimplerEggplant, SimplerTableCloth, SimplerPlateObjaverseLarger, SimplerSink, plus two 3 cm box cubes) and the backdrop photos download on first use from the eden_simpler dataset into simpler/ under the asset directory (SimplerObjectOptions.fetch_family_asset). See Assets.