DreamZero-DROID: DROID10 post-training, 1,000 updates

Public native DreamZero model-weight checkpoint, initialized from the complete official GEAR-Dreams/DreamZero-DROID checkpoint (revision 96ad344138c66e82536422432ad742f015784942). Full fine-tuning used 205 original DROID training episodes across 10 tasks and their original short instructions. No generated videos were used.

This is checkpoint 1,000, not a completed 3,000-step run. Training last logged step 1,045 before both Slurm allocations ended with NODE_FAIL on November 13, 2026. The latest saved checkpoint is step 1,000.

Contents

checkpoint-1000/ contains all 10 native safetensors model shards, the shard index, native model configuration, and the saved experiment configuration, normalization metadata and I/O provenance. All 2,146 model tensor names and shapes were structurally checked against official initialization. Every uploaded file is covered by the SHA256 manifest and remote upload verification.

DeepSpeed optimizer partitions, RNG state and trainer state are not uploaded. Complete optimizer/resume state remains in the local checkpoint. This repository provides model weights and configuration, not standalone optimizer-state recovery. No training videos, episode action files or raw observations are included.

The native checkpoint has not yet undergone a full final numerical audit, optimizer restoration or robot evaluation. Structural verification and upload hashes do not establish task success or convergence.

Training recipe

  • Official DreamZero source commit: ab790c198fbce33503358efbbd4187ce9a89adf3.
  • Full diffusion/action fine-tuning: approximately 16.484B trainable parameters; frozen native T5, CLIP and VAE components.
  • 8 B200 GPUs across two nodes; per-device batch 1, gradient accumulation 4, global batch 32.
  • Learning rate 1e-5; cosine schedule, 5% warmup over the requested 3,000 steps; weight decay 1e-5; Adam betas 0.95/0.999 and epsilon 1e-8.
  • BF16, native internal gradient checkpointing, DeepSpeed ZeRO-2 with CPU optimizer offload.
  • Source observations at 15 FPS; native video sampling at 5 FPS, up to 33 video frames, four action/video chunks, 24 actions per chunk and up to 96 target actions.
  • Native 176x320 single-view inputs and 352x640 three-camera collage. Physical-left source exterior2 is mapped to native exterior1 at bottom left; source exterior1 to native exterior2 at bottom right; wrist occupies the top.
  • Original RGB pixels, native transforms and original short instructions retained. Pixel-exact predecoded RGB cache, zero data-loader workers and an initial CPU readiness gate address decoder startup latency; these change loading/partition order without changing episode membership or model mathematics.

Download

from huggingface_hub import snapshot_download
root = snapshot_download("Cosmos-Data/dreamzero-droid10-1k")
checkpoint_dir = root + "/checkpoint-1000"

Use the official DreamZero repository and its native checkpoint loader. Experiment YAML paths reflect the originating cluster and require local path configuration. Official tokenizer/text/vision/VAE dependencies and upstream model terms remain applicable. This is not a Transformers AutoModel export.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Model tree for Cosmos-Data/dreamzero-droid10-1k

Finetuned
(1)
this model