· Research & technology· 日本語で読む
I-BFM Releases H200 Training Code and 166.560M-Step Checkpoint
The I-BFM repository provides H200 training code and a 166.560M-step checkpoint. Carry Box, Push Box and Kick Box tasks use the same checkpoint. The replay buffer for resuming training is about 210 GiB and is divided into eight shards.
Key facts
- The repository provides H200 training source code, a 166.560M-step checkpoint and run configurations for Carry Box, Push Box and Kick Box.
- The training instructions use eight H200 GPUs.
- The best policy is selected as the one with the lowest mean EMD across all 2,719 evaluation motions.
- The 166.560M-step replay buffer is provided in eight shards totaling about 210 GiB.
- The training source code carries over the upstream CC BY-NC 4.0 license.
What happened
The I-BFM GitHub repository has released training source code for H200, a selected 166.560M-step checkpoint, and run configurations for Carry Box, Push Box and Kick Box. Scripts for the three tasks use the same checkpoint. The instructions show training on eight H200 GPUs. The replay buffer at 166.560M steps is about 210 GiB and is provided in eight shards.
Why it matters
In addition to training code, the release includes the selection criteria based on 2,719 motions and the 166.560M-step replay buffer, making it easier for researchers to track not only policy execution but also training and evaluation conditions. However, the buffer is about 210 GiB, and loading it requires expanding an entire shard in memory, so use requires storage and computing resources.
Analysis
The release centers not only on distributing a completed policy but also on making training code, the selected checkpoint, task run configurations and the replay buffer available within the same project. The fact that all three tasks use the same checkpoint indicates a design in which a shared training result is tested through multiple task configurations, rather than distributing separate weights for each task. However, the repository description alone does not confirm performance on physical hardware or generalization across tasks.
Reproducibility has two aspects. The repository preserves training source code, input files, seeds and training parameters, and records the selection criterion as mean EMD across 2,719 motions, providing clues for tracking execution conditions and policy selection. On the other hand, it states that resuming without the replay buffer from the same step will produce results different from an uninterrupted run. Therefore, having access to the code and checkpoint is not the same as being able to reproduce the training process under exactly the same conditions.
The scale of the released data also affects operational requirements. The replay buffer totals about 210 GiB, is split into eight parts and requires the entire selected shard to be expanded in memory when loaded. The roughly 28 GB size of each shard indicates that even evaluating without handling all the data requires storage and memory on the user’s system. Together with the training procedure using eight H200 GPUs, this means using the released code requires substantial computing and data resources.
A question for further study is how mean EMD on evaluation motions corresponds to behavior during actual task execution. The release describes the policy selection criteria and how to run the interactive tasks, but does not report validation on physical robots, runtime safety or success rates for each task. The scope of reproducibility demonstrated by the repository should therefore be assessed separately from the additional validation required for real-world deployment.
Details
The repository includes training code and dependencies for running tasks, and describes a method for selecting the policy with the lowest mean EMD across all 2,719 evaluation motions. Copies of the released policy include weights and configuration, but not the optimizer or replay data. At the end of a normal training run, the replay buffer is written once to the final rolling checkpoint.
The replay buffer is available on Hugging Face as eight rank-specific shards. Each shard is about 28 GB on disk, and loading it expands the entire selected shard in memory. The repository explains that resuming without the replay buffer from the same step will produce results different from an uninterrupted run. Requirements include Linux, Python 3.10, uv, Git LFS and MuJoCo. The training source code carries over the upstream CC BY-NC 4.0 license.
Sources
Translated from the Japanese original: /news/i-bfm-h200-muke-gakushuu-koodo-to-166-560m-suteppu-no-chekku-de01cf. The Japanese version is authoritative if they differ.