The Policy Zoo¶

Eden ships published controllers that run as released, with no training: humanoid locomotion and motion tracking, skateboarding, dexterous tool use, insertion, and clients for server-hosted vision-language-action models. This page lists every policy in eden/extensions/policies/, what each one needs to run, and how eden inference --policy builds its world.

Published controllers running in Eden: SONIC, HUSKY, BeyondMimic, TWIST2, ProtoMotions, AMO, HOMIE, MetaMotivo and SimToolReal

Run a policy¶

eden inference --policy sonic --motion_file examples/projects/mimic/beyondmimic/sample/G1_gangnam_style_V01_bvh_60hz.npz --vis
eden inference --policy holosoma --base_command 0.5 0.0 0.0 --vis     # walk forward at 0.5 m/s
eden inference --policy beyondmimic_moves --clip_name B_DadDance --vis
eden inference --policy husky --rec                                   # record an mp4

Weights download on first use from the policy's Hugging Face dataset into the asset cache (policies/ under EDEN_ASSETS_DIR, by default ~/.cache/eden/assets), and are checked against the MD5 the policy pins. Policies with ONNX weights need the onnx extra: pip install 'eden-world[onnx]', or uv sync --inexact --extra onnx in a checkout.

--policy accepts the package name (holosoma), the registry key (holosoma_locomotion_policy), or the key without its _policy suffix. The flags that matter for a policy run:

Flag	What it does
`--policy NAME`	The policy to roll out. Mutually exclusive with `--checkpoint`.
`--task TASK`	Run the policy on a task's env instead of the policy's own world (below). Optional.
`--robot NAME`	The `eden.robots` class for a task-less run of a policy that does not pin its own robot (default `G1Dofs29`). Ignored when the policy pins one, and with `--task`.
`--entity_name NAME`	The scene entity the policy controls (default `robot`).
`--motion_file PATH`	The reference motion for a motion-tracking policy. Required when the policy takes one.
`--clip_name CLIP`	The clip for a per-clip policy (`beyondmimic_moves`); ignored by the others.
`--base_command VX VY WZ`	A constant command for a command-driven policy (default: stand). Padded or truncated to the policy's command width.
`--steps N`	Steps to run (default `1000`). `--episodes` is not available with `--policy`.
`--num_envs N`	Parallel envs, with `--task` only: a task-less run builds one env.
`--vis`, `--rec`, `--async`, `--cpu`	As for a checkpoint replay (Training Runs).

Task-less runs and `--task` runs¶

Without --task, Eden builds the policy's own world from its spec: one env, an empty scene holding the robot robot_options() returns plus every entity scene_entity_options() declares (the skateboard, the tool and its table), the policy's env_options() (its control rate and solver settings), and its observation, action and command terms. This is the run the policy was ported for, and the one to start with.

With --task, the policy runs on that task's env, and the policy's command terms are merged into the task's. The task must provide what the policy reads: the entities (checked at construction), and, for a policy that reads the env's observations, the exact observation group it was built for. A policy that observes its command through the task's own command terms ignores --base_command, and passing a non-zero one is an error.

The shipped policies¶

Humanoid locomotion¶

`--policy`	What it is	Robot	Needs	Control
`holosoma`	Holosoma velocity-tracking locomotion (ONNX); also the safety base `eden deploy` preloads on a G1	`G1Dofs29` at Holosoma's default pose	`--base_command VX VY WZ`; `onnx` extra	50 Hz
`homie`	OpenHomie walk/stand tracker for the lower body (walk and stand torch checkpoints)	`G1Dofs29` with HOMIE's pose and gains	`--base_command`: a 4-D `[vx, vy, wz, height_delta]` command, padded with zeros	50 Hz
`amo`	AMO adaptive whole-body control (TorchScript)	`G1Dofs29NoWristDof` (23 DOF)	Nothing: its command term resamples its own commands	50 Hz
`husky`	HUSKY humanoid skateboarding (ONNX)	`HuskyG1` (23 DOF) plus a skateboard	`--base_command`: a 2-D `[lin_vel_x, relative_heading]`; `onnx` extra	50 Hz
`limx_oli_loco`	Velocity-tracking loco/stand controller for LimX Oli (rsl_rl actor)	`Oli`	No released weights: a local `velocity_tracking_flat` checkpoint trained for Oli (see Local checkpoints)	the env default

Humanoid motion tracking¶

These track a reference motion. Each takes --motion_file (an Eden motion .npz, see Motion Tracking) except where noted.

`--policy`	What it is	Robot	Needs	Control
`sonic`	GEAR-SONIC universal-token whole-body controller (encoder + decoder ONNX)	`G1Dofs29` at SONIC's standing pose	`--motion_file`; `onnx` extra	50 Hz
`twist2`	TWIST2 whole-body tracker (ONNX)	`G1Dofs29`, knees bent	`--motion_file`; `onnx` extra	50 Hz
`humanoid_gpt`	Humanoid-GPT tracker, Transformer by default or MLP (ONNX)	`--robot` (default `G1Dofs29`)	`--motion_file` (a 50 Hz motion); `onnx` extra	50 Hz
`protomotionv3`	ProtoMotions v3 `g1-bones-deploy` tracker (torch actor from the training checkpoint)	`G1Dofs29`	`--motion_file`	50 Hz
`omnixtreme`	OmniXtreme flow-matching base policy plus residual correction (three ONNX graphs)	`OmniXtreme_G1`	`--motion_file`; `onnx` extra	50 Hz
`gmt`	GMT general motion tracking (TorchScript)	`G1Dofs29NoWristDof` (23 DOF)	`--motion_file`: GMT's own `.pkl` format (samples in `examples/policies/gmt/motions/`)	50 Hz
`beyondmimic_moves`	60 BeyondMimic tracking policies from the `exptech/g1-moves` dataset, one per clip	`G1Dofs29` at the g1-moves pose	`--clip_name` (default `B_DadDance`); the clip's checkpoint and motion download on first use	50 Hz
`metamotivo`	Meta Motivo (FB-CPR) behavior foundation model for an SMPL humanoid; behavior set by a context vector `z`	`SmplHumanoid`	The `metamotivo` package (`pip install "metamotivo[huggingface]@git+https://github.com/facebookresearch/metamotivo.git"`); the model loads from Hugging Face (`facebook/metamotivo-S-1` by default)	30 Hz

The sample motion examples/projects/mimic/beyondmimic/sample/G1_gangnam_style_V01_bvh_60hz.npz is what the sonic, humanoid_gpt and omnixtreme examples default to; examples/policies/twist2/resources/greeting.npz is TWIST2's.

Manipulation¶

`--policy`	What it is	Robot	Needs	Control
`simtoolreal`	SimToolReal object-centric tool manipulation (rl_games LSTM actor reimplemented in torch)	`KukaSharpa` (KUKA iiwa14 arm + SharPa hand, 29 DOF) plus a table, a DexToolBench tool and a goal ghost	Nothing extra; the tool defaults to `long_screwdriver`	60 Hz
`omnireset`	UWLab OmniReset state-based insertion expert	`OmniResetUR5e2f85` (UR5e + Robotiq 2F-85) plus the object pair	The `peg` pair is the one released expert	10 Hz

Server-backed clients (VLAs and world models)¶

These policies run inference on a separate server and only exchange observations and actions with it. They need the policy-client extra (pip install 'eden-world[policy-client]'), a running server, and a manipulation scene with the cameras their setup reads; see Remote Policy Clients.

`--policy`	Server	Transport	Default address	Setups (`policy_setup`)
`pi`	OpenPI (pi0, pi0-FAST, pi0.5)	`openpi_websocket`	`localhost:8000`	`widowx_bridge` (default), `google_robot`, `libero`, `droid`, `droid_jointpos`, `aloha`
`gr00t`	NVIDIA Isaac-GR00T	`zmq_msgpack`	`localhost:5555`	`simpler_env_widowx` (default), `libero_sim`, `oxe_droid_relative_eef_relative_joint`
`galaxeavla`	GalaxeaVLA (G0.5)	`galaxea_websocket`	`localhost:8765`	`galaxea_r1lite` (default), `galaxea_r1pro`, `droid_franka`
`cosmos3`	Cosmos3-Nano-Policy-DROID (world action model)	`openpi_websocket`	`localhost:8000`	DROID only
`molmoact2`	MolmoAct2 DROID server (absolute joint control)	`http_json_numpy`	`localhost:8000`	DROID only
`octo`	Octo, through an HTTP server	`http_json`	`localhost:9010`	WidowX only

eden inference --policy <client> constructs the client with its defaults (the address, setup and no instruction above), so to choose a server, a setup or a language instruction, drive the client from Python as Remote Policy Clients shows.

Local checkpoints¶

Every policy's file locations are class constants, not options. To run your own weights in place of the released ones, call use_local_checkpoint on the class before constructing it; it also clears the MD5 pinned to the released file, which your file would fail:

from eden.extensions.policies.limx_oli_loco.policy import OliVelocityLocoPolicy

OliVelocityLocoPolicy.use_local_checkpoint("logs/velocity_tracking_flat/<run>/model_3000.pt")

A policy that loads several files names the one to redirect: SonicPolicy.use_local_checkpoint(path, checkpoint_attr="decoder_checkpoint_path"). On the command line, eden deploy --policy NAME --policy_checkpoint NAME=PATH does the same for a deployment.

To fetch every policy's files from a mirror (an offline cluster, a shared read-only cache), export EDEN_ASSETS_DIR=/data/mirror: policies then cache under /data/mirror/policies/.

Policies on hardware and in MuJoCo¶

eden deploy --policy NAME loads any of these as a switchable slot on a real robot or, with --mujoco, in MuJoCo. On a G1, Holosoma is preloaded as the safety base slot. Which policies can share one deployment, and how slots switch, is in Deployment. examples/policies/policy_switch/main.py shows the sim-side counterpart: Holosoma and SONIC hot-swapping on one G1.

Example scripts¶

Most policies also have a standalone script under examples/policies/<name>/main.py that builds the same world and adds policy-specific controls (a keyboard command, a clip picker, recording): amo, beyondmimic_moves, gmt, holosoma, homie, humanoid_gpt, husky, metamotivo, omnireset, omnixtreme, protomotionv3, simtoolreal, sonic and twist2. Each script's header gives the command to run it.