Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
512
4.1k

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

RoboDojo put_bottles_into_dustbin on Isaac Lab

One RoboDojo evaluation layout (layout set 0, layout 1) of put_bottles_into_dustbin, ported to a client-defined Isaac Lab 2 environment. Two ARX X5 arms must put every bottle on the table into the dustbin.

Instruction: Pick up the bottles and throw them into the dustbin, using handover when needed.

Earlier π0.5 rollout

π0.5 rollout: head camera, left and right wrist cameras

Download the rollout video.

The retained video was recorded before the native mass/material repair. RoboDojo's released π0.5 checkpoint completed that earlier revision of this layout:

Outcome success
Score 1.00
Steps 312 of 700 (25 Hz control; 12.5 s of simulated time)
Policy calls 7 (50-action chunks, each executed in full)
Median env.step / policy call 0.31 s / 155 ms

The checkpoint is ckpt/RoboDojo/Pi_05/RoboDojo-sim-arx_x5-joint-0/59999 from the Hugging Face dataset RoboDojo-Benchmark/RoboDojo at revision 63fcc045f2e9, served with the vendored openpi of XPolicyLab 5d0be149 (pi05_base_aloha_full_sim_arx-x5_seed_0). Policy inputs are the three 640×480 cameras, the 14-value joint and gripper state and the instruction. The video lays out the head camera with the left and right wrist cameras at the 25 Hz control rate.

Success and score come from the ported RoboDojo evaluator running in the client. They are caller-reported, not platform-verified official scores, and one episode is not a success rate.

Run it

from simulator import load_env

env = load_env("pufanyi/robodojo-put-bottles-into-dustbin", realtime=False, trust_remote_code=True)
with env.start():
    observation = env.observe()  # vision, state, instruction and outcome
    # 14-value absolute joint and normalized gripper targets per step
    env.act({"left_arm": [0.0] * 6, "left_ee": [1.0], "right_arm": [0.0] * 6, "right_ee": [1.0]})

Run it with realtime=False: the evaluator observes every step. env.observe()["outcome"] holds the step, score, success and time-out; a finished episode resets automatically and its result stays in observe()["ended"].

This layout runs with Isaac Lab's Fabric buffers (sim.use_fabric = True); RoboDojo itself uses use_fabric = False.

The prepared scene includes effective rigid-body mass and collision material values read from the original runtime. Startup restores these before settling, preserving the original inertia. This layout has four 0.15 kg bottles with static and dynamic friction 0.5. Its scene and metadata are checksum-pinned in config.yaml.

Downloads last month
383