Evaluation sets of the Praxis-1 robot policies. One folder per evaluation. The models are in the Praxis-1 collection; the reference code is in runwayml/praxis-1-oss.

Folder	Evaluation	Sets
`autoregressive/`	Video-only autoregressive rollout: does the imagined video follow the prompt?	`droid`, `yam_runway`, `yam_molmoact2`

`autoregressive/`: video-only autoregressive prompt adherence

A model gets one real first frame and a task prompt. It rolls its video out on its own predictions for 12.6 s (189 frames at 15 fps, no ground-truth frames or actions after frame 0). A vision-language judge then scores the video: did the robot act on the named object, bring it to the named destination, and perform the named action?

The headline metric is full adherence = target object acted on AND destination used AND action correct. It is the mean over seeds per prompt, then the mean over prompts, with a 95 % bootstrap interval over prompts. Two models are compared paired, on the same (prompt, seed).

The judge is GPT-6 Astra with a fixed system prompt. It sees 12 evenly spaced frames and the end state. Against 80 human labels its agreement on full adherence is kappa 0.87.

`autoregressive/droid`


Robot	single-arm Franka (DROID rig)
Views	exterior + wrist, one 832x224 canvas (two 416x224 tiles, exterior left)
Scenes	8 first frames from 6 DROID 1.0.1 episodes of the val split, 4 labs (AUTOLab, CLVR, GuptaLab, IPRL)
Prompts	35, each a pick-and-place instruction with look-alike distractors
Rollouts	104 = 35 prompts x 2 to 4 seeds
Features	colour (easy, medium), type (easy, medium), spatial (easy), count (easy, hard)
Actions	`place_on` (54), `place_in` (50)

Each scene is a real DROID first frame. The objects of the task (target, look-alike distractors, receptacles) were rendered at the calibrated camera poses and inserted into both views by an image model. Each insertion passed automatic geometry and appearance checks and a manual review. The objects, their positions and the prompts are new: no model has seen these scenes, only the DROID backgrounds.

Files

frames/<scene>.png: the first frame of each scene (lossless).
frames_ext2/<scene>.png: the second exterior view of each scene at the same 416x224 tile size (lossless): the episode's real second exterior frame with the same objects inserted. Praxis-1 does not use it; models with a second exterior camera input (the baselines below) get it there.
scenes.jsonl: one row per scene: scene, frame, sha256 of the PNG, frame_ext2, sha256_ext2 (the second exterior view), droid_episode (DROID episode id), lab, inserted_objects, views, objects (with rests_in and state), the identical twins, and exterior_center_px: the pixel centre of each object in the 416x224 exterior tile, which the judge uses to tell identical instances apart.
rollouts.jsonl: one row per rollout: id, scene, prompt, seed, target_object, destination, action, distractors (each with the attribute it differs_by), destination_distractor, feature, level, and for the spatial and count features relation, relation_applies_to, required_count, count_mode, n_present, target_instances (the identical instances that count) and count_filter.

Scene ids read <feature><level><index>_<lab>_<date>: co colour, ty type, sp spatial, ct count; e easy, m medium, h hard.

Reference results (full adherence, %, 95 % bootstrap interval over prompts; the real frame-0 robot state; each baseline in its joint video + action mode with its released default sampler, given the second exterior view of frames_ext2/; Praxis-1 with the eval sampler: 10 Euler steps, text guidance 3 with a negative prompt, flow shift 8):

Model	Full clip (12.6 s)	Within 11.9 s
Praxis-1 (`praxis-1-droid`)	52.9 [38.6, 67.1]	46.4 [31.4, 61.4]
FLUX 3 Action DROID	48.6 [35.0, 62.9]	42.1 [27.9, 57.1]
DreamZero-DROID	42.1 [30.0, 55.0]	37.9 [25.7, 51.4]
Cosmos3-Nano-Policy-DROID	22.9 [11.4, 35.7]	17.9 [7.1, 30.7]
OpenWAM-alpha (Foundation)	1.4 [0.0, 4.3]	0.0 [0.0, 0.0]

OpenWAM-alpha has no second exterior input. With a black second exterior view instead, the baselines score lower: FLUX 3 Action 38.6 [24.3, 53.6], DreamZero-DROID 26.4 [12.9, 40.7], Cosmos3-Nano-Policy-DROID 23.6 [11.4, 37.1] (within 11.9 s: 36.4, 19.3, 22.9).

11.9 s is the time by which the fastest of these baselines (DreamZero) completes 95 % of its successes. The full clip favours models whose video stays stable; the 11.9 s cut-off removes that advantage.

`autoregressive/yam_runway` and `autoregressive/yam_molmoact2`

Two sets for the bimanual YAM rig (two I2RT YAM arms, a top camera and a camera on each wrist). They differ in the workcell and the top camera:

	`yam_runway`	`yam_molmoact2`
Backgrounds	the Runway YAM booth: an enclosed workcell, high top camera	the MolmoAct2 BimanualYAM rig: an open wooden table, low top camera
Source episodes	7: 6 from ABC-130K, 1 from a Runway capture	8 from 5 MolmoAct2-BimanualYAM sessions (untangle task)
Views	top, left wrist, right wrist; one 1248x256 canvas (three 416x256 tiles)	top, left wrist, right wrist; one 960x176 canvas (three 320x176 tiles), plus each view at 640x360 in `frames_640/`
Scenes	56	36
Prompts	262	168
Rollouts	886	564
Seeds	2 per easy prompt, 4 per medium or hard prompt	same
Features	colour, type, size, spatial, count, containment; each at easy, medium and hard	same

Each set is in the training domain of different models. The Runway booth is in Praxis-1's training data. MolmoAct2-BimanualYAM is in the training data of Praxis-1, DreamZero-YAM and Flex-π. The inserted objects, their positions and the prompts are new in both sets: no model has seen these scenes, only the backgrounds. The sets differ in scenes and camera, so compare models within a set.

Each scene row gives the real robot state at the first frame of its source episode (state0, 14-D joints, gripper 1 = open), for models that need a proprioceptive input.

Files (per set): frames/<scene>.png, scenes.jsonl (as for droid, plus source, source_episode or source_repo / source_episode_index, state0, state0_layout; yam_molmoact2 also lists frames_640) and rollouts.jsonl (as for droid, plus container for the containment feature). The scene rows also list the objects (with rests_in and state), the identical twins, and top_center_px: the pixel centre of each object in the 416x256 top tile, which the judge uses to tell identical instances apart. Scene ids read <feature><level><index>_<source episode>: co colour, ty type, sz size, sp spatial, ct count, cs containment; e easy, m medium, h hard.

yam_runway keeps four prompts with a doubled word ("Put the small small box ..."), from a noun error in the object catalog. They are kept as they are so that every result on the set stays comparable.

Reference results (full adherence, %, 95 % bootstrap interval over prompts; the baselines in their joint video + action mode with their released default sampler and the real frame-0 state; Praxis-1 with the eval sampler: 10 Euler steps, text guidance 3 with a negative prompt, flow shift 8):

Model	`yam_runway`	`yam_molmoact2`
Praxis-1 (`praxis-1-base`)	48.0 [43.5, 52.5]	55.2 [49.0, 61.6]
DreamZero-YAM	0.7 [0.2, 1.2]	14.7 [10.9, 18.9]
Flex-π (`flexpi-pretrain-rgb-preview`)	0.0 [0.0, 0.0]	0.0 [0.0, 0.0]

The results of the released bimanual checkpoint praxis-1-yam on both sets will be added here.

Usage

from huggingface_hub import snapshot_download

path = snapshot_download("runwayml/praxis-1-eval", repo_type="dataset",
                         allow_patterns=["autoregressive/droid/*"])  # or yam_runway, yam_molmoact2

Load the models with runwayml/praxis-1-oss.

Provenance and licence

droid: the backgrounds are frames of DROID 1.0.1 (Khazatsky et al., 2024), CC BY 4.0.
yam_runway: six backgrounds are frames of ABC-130K (Apache-2.0); one is a Runway capture.
yam_molmoact2: the backgrounds are frames of MolmoAct2-BimanualYAM (Apache-2.0).

The inserted objects, the prompts and the Runway capture are by Runway. This dataset is released under the Apache License 2.0. The DROID backgrounds remain subject to the attribution terms of CC BY 4.0; please credit Praxis-1 and the source dataset of each set you use.

Citation

@article{praxis1,
  title  = {Praxis-1},
  author = {Runway},
  year   = {2026}
}

@inproceedings{khazatsky2024droid,
  title     = {{DROID}: A Large-Scale In-The-Wild Robot Manipulation Dataset},
  author    = {Khazatsky, Alexander and Pertsch, Karl and others},
  booktitle = {Robotics: Science and Systems},
  year      = {2024}
}