Datasets:

---

BitRobot

/

HIW-500-LeRobot

like 21

Follow

BitRobot 58

Unfortunately the whole body controller that was use to record this is proprietary, but in terms of open-source wbc i'm you can replicate this closed-source controller's behaviour with groot-wbc (https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/unitree_g1/gr00t_locomotion.py): in LeRobot we also support holosoma, and we will be porting sonic wbc in the coming weeks (PR here https://github.com/huggingface/lerobot/pull/3827)

No idea what gripper is being used, i'll try asking the authors but maybe they can answer here.

The observations we support can be visualized here: https://huggingface.co/spaces/lerobot/visualize_dataset?path=%2FBitRobot%2FHIW-500-LeRobot%2Fepisode_0. In short, yes, the original dataset BitRobot/HIW-500 contains more (IR camera, odometry etc, these are not included as we do not support them in LeRobot yet)

To deploy on real G1 follow this guide https://huggingface.co/docs/lerobot/en/unitree_g1. For current wbc compatibility, as it currently stands I would recommend replaying one episode on the G1 scaling pivot_vx, vy, height and vyaw in sim (https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) until sim and 3d replay match, then simply training on them. Ideally if you wanna wait until SONIC is ported then training would be much simpler, full robot pose is included in the observations so you could just output robot state and let sonic approximate it directly which should be easier.

Thanks for the previous clarification about the WBC deployment path. The Groot / Sonic-style WBC replay side is working well on our end now. The remaining blocker is the gripper semantics.

We are replaying HIW-500-LeRobot episodes on a Unitree G1 with a Dex1-1 parallel-jaw gripper, and we found that the gripper channels do not look like direct jaw-position targets.

In both action and observation.state.wbc, the last 4 dimensions are:

[19] left_trigger
[20] left_squeeze
[21] right_trigger
[22] right_squeeze

From meta/stats.json, trigger is roughly in [0, 10], while squeeze is in [0, 1]. Also, for the episodes we checked, the gripper part of action and observation.state.wbc is bit-identical, so it looks like observation.state.wbc is echoing the command rather than providing independent jaw-width feedback.

The ambiguity is that trigger does not behave like a continuous open/close position. It stays near 10 for most frames and only briefly dips toward 0. In the videos, the same value trigger = 10 can correspond to both an open and a closed gripper depending on history.

For example, in episode 12, hang keys on a hook, right gripper:

time 12s: right_trigger = 10, gripper open
time 16s: right_trigger = 0,  gripper closed / grasping
time 22s: right_trigger = 10, gripper still closed / holding keys

So trigger = 10 cannot simply mean “fully open”. Similar behavior appears in other episodes. If we treat downward pulses of trigger as toggle events, where each pull toward 0 flips the latched gripper state open/closed, the predicted open/close state matches the videos much better.

Could you clarify the intended semantics?

1. Is trigger supposed to be a continuous jaw opening / position target, or is it a teleoperation trigger event that toggles a latched open/closed gripper state?

2. If it is stateful/latched, what is the initial gripper state at the start of an episode? We are currently assuming open.

3. What does squeeze represent for Dex1-1? Is it an extra grip-force / hard-close command, or something else? Since it is near zero in most frames, is it safe to ignore for basic open/close reconstruction?

4. Is there any documented mapping from (trigger, squeeze) to the actual Dex1-1 command, such as jaw position, velocity, force, or open/close latch state?

5. Are these gripper fields synchronized to the same timestamp as the camera frames, or is there a known latency between the low-dimensional stream and the videos?

This matters because if we map trigger directly to jaw width, the simulated gripper appears open for most of the episode, which disagrees with the real wrist/head videos. If the correct interpretation is a toggle/latch event, we should reconstruct the gripper state before training or replaying policies.

Thanks again for the help!

Would you mind moving this discussion to the LeRobot discord? Thank you! https://discord.gg/3MSyYeG2j

Sure, thank you!

I’ve also created an open-source HIW-500 controller/replay implementation. Hopefully it will be useful to anyone who finds this discussion later:

https://github.com/Hxxxz0/HIW-500-controoler

amazing work

have you had any luck training on this?

yes

do you have any tips for training? i forked your repo here https://github.com/nepyope/HIW-500-controller-LeRobot, we ship gr00t inside lerobot and i'm currently trying to train. The controller code was really useful especially for the gripper, seems like groot works decently but i probably need some more finetuning? here's the run i did on "put the pillow from the floor to the sofa" tasks https://wandb.ai/tino/hiw500-pillow/runs/t8bb38ih?nw=nwusertino. It's about 2k episodes or ~20 hours of data, did you do any finetuning on your setup to get it to work?

Glad to see the controller code was useful, and great work getting GR00T running on your fork! I’ve also been experimenting with different models on HIW-500, including π0, but I’m still in the early stages of exploring the training setup. Unfortunately, I don’t currently have a compatible gripper-equipped robot for deployment and real-world evaluation, lol.
My training loss is actually in a similar range to yours. From what I’ve seen so far, though, I feel that GR00T may work better for this task with some additional task-specific finetuning. I’d be very happy to share any useful findings as I continue experimenting, and please feel free to share your results or any tricks you discover as well!

comparison_sim.mp4