Ujere/phasor-m7-robocasa365

Phasor_m7 — RoboCasa365

Not affiliated with Xiaomi. This is a fine-tune of Xiaomi's real, open-weight (Apache 2.0) Xiaomi-Robotics-1-RoboCasa365 checkpoint. The base model (Qwen3-VL-4B VLM + 36-layer DiT action head) is kept entirely frozen — verified on every training run via an exact weight-sum match before/after. On top of it we add:

• Rank-4 LoRA adapters on the VLM's language_model attention/MLP layers only
• Phasor's own sparse end-to-end memory core (Hopfield network + HGRN gate + adaptive gate) applied to pooled VLM hidden states, as an additive consolidation-loss regularizer during training

Full training/eval code, tests, and raw per-episode evidence: https://github.com/Phasor-inc/minimalist_m7

Real, final results (RoboCasa365 target50 protocol, 2500 episodes)

Run	atomic_seen	composite_seen	composite_unseen	Overall
Baseline (frozen Xiaomi-Robotics-1-RoboCasa365)	79.44%	57.63%	30.75%	56.88%
This checkpoint (LoRA + M7)	78.89%	55.00%	29.88%	55.56%

Honest read: this checkpoint is essentially tied with the frozen baseline on atomic_seen and modestly behind on composite_seen/composite_unseen/overall. See the GitHub repo for the matching LoRA-only ablation (isolating the memory core's specific contribution) and full per-episode results.

Training config

• Base: XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa365, initialized from Qwen/Qwen3-VL-4B-Instruct
• Optimizer: AdamW, lr 1e-4, no scheduler
• LoRA: rank 4, alpha 8.0, zero-init B
• Batch size 48, action length 30, 220 training steps
• M7 memory core: consolidation weight 0.1, adapter target dim 512, memory size 256
• Training data: real RoboCasa365 data (NVIDIA PhysicalAI-Robotics-Manipulation-Kitchen-Demos), restricted to the 34 official atomic_seen+composite_seen tasks; composite_unseen correctly held out of training

Requirements

The reference environment uses:

Python 3.11
PyTorch 2.8.0
Transformers 4.57.1
FlashAttention 2

This checkpoint uses the same custom Transformers model/processor code and RoboCasa365 action normalization statistics as the base Xiaomi checkpoint, loadable via AutoModel.from_pretrained(..., trust_remote_code=True).