Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
label
string
robot
string
domain_id
int64
example
string
native_domain
string
cross_embodiment
bool
camera
string
frames
int64
fps
int64
seed
int64
start_image
string
video_frames
int64
width
int64
height
int64
duration_s
float64
gen_seconds
float64
prompt
string
created
float64
file_name
string
actions_file
string
initial_image
string
hardware
string
20261114-142300-imp61
Mixed 028 · first test (61 fr)
agibotworld
15
agibot_mixed_028_bulk_seed2026
agibotworld
false
recorded
61
30
2,026
agibot_mixed_028
61
640
352
2.03
164
The scene opens in a continuous landscape shot from a high oblique overhead angle, sharply focused across a glossy white kitchen countertop with faint gray streaking, soft reflections, and diffuse cool-neutral indoor light that casts gentle moving shadows and gives the workspace a clinical, controlled demonstration qua...
1,794,646,380
runs/20261114-142300-imp61/video.mp4
runs/20261114-142300-imp61/actions.json
runs/20261114-142300-imp61/initial.png
NVIDIA Jetson AGX Orin 64GB
20261114-163100-imp901
Mixed 028 · first test (901 fr)
agibotworld
15
agibot_mixed_028_bulk_seed2026
agibotworld
false
recorded
901
30
2,026
agibot_mixed_028
901
640
352
30.03
2,105
The scene opens in a continuous landscape shot from a high oblique overhead angle, sharply focused across a glossy white kitchen countertop with faint gray streaking, soft reflections, and diffuse cool-neutral indoor light that casts gentle moving shadows and gives the workspace a clinical, controlled demonstration qua...
1,794,654,060
runs/20261114-163100-imp901/video.mp4
runs/20261114-163100-imp901/actions.json
runs/20261114-163100-imp901/initial.png
NVIDIA Jetson AGX Orin 64GB
20261114-164953-bf79
YAM xdof · dashboard test
xdof_yam
16
yam_019_bulk_seed2026
xdof_yam
false
static
61
30
2,026
yam_019
61
640
480
2.03
222.1
The scene opens directly into an overhead wide shot from a high, nearly vertical angle in landscape framing, sharply focused across a compact indoor corner with bright, even overhead lighting, minimal shadows, moderate contrast, and a cool neutral palette of white, gray, black, pale wood, lavender-purple, pale green, a...
1,794,655,193.160873
runs/20261114-164953-bf79/video.mp4
runs/20261114-164953-bf79/actions.json
runs/20261114-164953-bf79/initial.png
NVIDIA Jetson AGX Orin 64GB
20261114-193433-6c7c
molmoact2_yam
16
yam_019_bulk_seed2026
xdof_yam
true
recorded
121
30
2,026
yam_019
121
640
480
4.03
372.8
The scene opens directly into an overhead wide shot from a high, nearly vertical angle in landscape framing, sharply focused across a compact indoor corner with bright, even overhead lighting, minimal shadows, moderate contrast, and a cool neutral palette of white, gray, black, pale wood, lavender-purple, pale green, a...
1,794,665,073.177316
runs/20261114-193433-6c7c/video.mp4
runs/20261114-193433-6c7c/actions.json
runs/20261114-193433-6c7c/initial.png
NVIDIA Jetson AGX Orin 64GB

Cosmos3 Bimanual on Jetson AGX Orin

Action-conditioned robot videos generated with nvidia/Cosmos3-Nano-Sim-Bimanual running on-device on an NVIDIA Jetson AGX Orin 64GB (vLLM-Omni vllm/vllm-omni:cosmos3, arm64).

Each run starts from one image and a sequence of bimanual robot actions (and optional camera motion); the world model renders how the scene evolves under those actions.

4 runs · 38 s of video · robots: AgiBot (gear gripper), YAM (MolmoAct2), YAM (xdof)

LeRobot v3.0 versions (one per robot family, for LeRobotDataset): agibot · yam

Runs

Run Robot Camera Frames Length Resolution Generation time Action source
20261114-142300-imp61 AgiBot (gear gripper) recorded 61 2.03 s 640x352 2.7 min agibot_mixed_028_bulk_seed2026
20261114-163100-imp901 AgiBot (gear gripper) recorded 901 30.03 s 640x352 35.1 min agibot_mixed_028_bulk_seed2026
20261114-164953-bf79 YAM (xdof) static 61 2.03 s 640x480 3.7 min yam_019_bulk_seed2026
20261114-193433-6c7c YAM (MolmoAct2) (cross-embodiment from xdof_yam) recorded 121 4.03 s 640x480 6.2 min yam_019_bulk_seed2026

Layout

metadata.jsonl              one row per run (settings, timing, file paths)
runs/<id>/video.mp4         generated video (30 fps)
runs/<id>/actions.json      exact actions sent: {action_space, domain_names[T], action[T][59]}
runs/<id>/meta.json         generation settings and timing
runs/<id>/initial.png       conditioning frame
runs/<id>/filmstrip.jpg     8 evenly spaced frames

Actions

actions.json holds the model's normalized unified_v1 action rows (59-D, one per generated frame after the first), exactly as sent to the server:

Slice Meaning
0:9 ego / camera pose: translation xyz + rot6d
9:18 right wrist pose: translation xyz + rot6d
18:33 right fingertips (zero / masked for gripper robots)
33 right gripper opening (−1 = closed)
34:43 left wrist pose
43:58 left fingertips (zero / masked)
58 left gripper opening (−1 = closed)

domain_names[t] is the embodiment of row t; camera_pose rows move only the camera.

  • camera = recorded: the trajectory's own camera motion.
  • camera = static: ego slots held at the "no motion" value on every row.
  • cross_embodiment = true: actions recorded on native_domain were sent labelled as robot.

Generation

  • Model: nvidia/Cosmos3-Nano-Sim-Bimanual (distilled, fixed 4-step sampler), deploy config cosmos3_nano_sim_bimanual_windowed.yaml (64-latent-frame rolling cache, 8 attention sinks).
  • Hardware: Jetson AGX Orin 64GB (Ampere sm_87; not an officially listed target). About 0.4 frames/s, i.e. ~35 min for a 30 s clip.
  • Start images, prompts and trajectories come from the example assets of the model repository.

License

Generated with a model released under the OpenMDW 1.1 license; see the model card for terms.

Downloads last month
99