microsoft

/

rho-yam-box

Like 8

Follow

Microsoft 22.8k

This PR fixes https://github.com/microsoft/rhobotics/issues/54 for this checkpoint.

Problem

policy.json declares "type": "phi4mm_tactile" (with "name": "rhoalpha"), a policy type that the current rhobotics code no longer registers. Evaluating or serving this checkpoint directly fails while its config is read:

draccus.utils.DecodingError: Couldn't find a choice class for 'phi4mm_tactile' in <class 'rho.policies.base.PolicyConfig'>

Fine-tuning from this checkpoint is unaffected, because training builds the policy from the training YAML and loads only the weights. Checkpoints fine-tuned from it are also unaffected, because they store their own "type": "rho" config.

Change

Two metadata files change; the weight files do not.

policy.json is converted to the Rho format that microsoft/rho-fr3-duo already uses:

• type and name are set to rho.
• 48 RhoAlpha-only settings are removed (listed below). RhoConfig has none of these fields, and the loader already discarded them.
• The name key inside lr_scheduler is removed; it duplicates type.
• "pretrained_repo_id": "microsoft/rho-base" is added. This is RhoConfig's default, so loading behaves the same either way; it is included so that the file has the same fields as rho-fr3-duo's.

All other values are unchanged. The weights already use the Rho layout: all 964 tensors match microsoft/rho-base in name, shape and dtype.

model.safetensors.index.json loses 651 model.text_output_model.* entries from its metadata. They are aliases left over from RhoAlpha's text-output head, each pointing to a tensor that is stored under its model.flow_model.* name. The rhobotics loader currently restores these aliases and then drops them again, logging "Dropped 651 non-Rho keys while converting RhoAlpha weights", so loading depends on RhoAlpha compatibility code. The weight map is unchanged. The only other edit is the trailing comma after rho_version, which is now the last metadata entry. No other Rho checkpoint (rho-base, rho-fr3-duo, rho-libero, rho-roboeval) has such aliases.

Verification

Tested with rhobotics main at fd9cc54:

• load_configs_from_checkpoint reads the config as RhoConfig. The decoded config matches rho-fr3-duo's except for num_flow_samples (1 vs. 8), which only affects training.
• Following the serving code path (EvalConfig, make_policy, load_from_pretrained, PolicyInterface.get_action_chunk) with flash_attention_2, all weights load under strict key checking, and the policy returns finite (1, 50, 20) action chunks that are identical across repeated calls.
• With the new index, loading no longer drops any keys, and the action chunks are bit-identical to those produced with the original index for the same inputs and noise.

Not in scope

This checkpoint's stats.json and data_config.json are copies of rho-base's. That is tracked separately in https://github.com/microsoft/rhobotics/issues/53. Until it is fixed, actions decoded with the shipped statistics are not correctly scaled; for example, gripper values fall outside [0, 1].