JAM RoboTwin Command-Delta Small β 100 epochs, EMA
Latest model: robotwin_command_delta_small_100e_20261111/last.ckpt, completed 100,000 training batches (100 Γ 1,000). This artifact contains inference-only EMA weights and complete model state/statistics/tokenizer/backbone configuration metadata. All 761 EMA tensor/buffer byte sequences equal the original checkpoint. See checkpoint_manifest.json for SHA256 and provenance.
Training population of this published checkpoint: Clean50 + selected Random50 per task (5,000 episodes), using the original experiment sampler. It was NOT trained on the new full Clean50 + Random500 50:50 mixture. The updated GitHub handover provides a separate recipe for that next training run.
- Small DINOv3, embedding 384; projector hidden 768.
- Predictor depth 6, heads 24 Γ dim_head 64, MLP 2048.
- Deterministic Actor depth 4, hidden 384, heads 6; bidirectional futures; command-delta targets with actual last-issued-command reference.
- MSE pose objective; action/current weights 64/6; last 4 encoder blocks fine-tuned.
- 3 RGB views + instruction, context 12, future 24 Γ 20ms targets, replan 24.
- Future command-delta outputs are decoded using the stored future normalization and current issued-command reference. Use the corresponding policy to initialize/reset/update that reference; do not execute deltas as absolute coordinates.
Code/setup/training/re-rendering/eval handover: https://github.com/Mengmouxu/jam-robotwin-eval . Start with docs/handover/AGENT_RUNBOOK.md; use scripts/run_eval.sh for this checkpoint. The historical E50 requires scripts/run_published_eval.sh and code under runtimes/small_abs_pose_e50.
The previous absolute-pose E50 remains available at revision 2e60207e69655f8dac3497fdaad4322a3634cba4. Loading incompatible historical checkpoints through new code is not supported.
Validation: original training runtime + original checkpoint vs handed-over current runtime + exported checkpoint, CPU FP32 fixed RGB/policy/native executor boundary replay: 44 arrays bitwise equal. This does not assert physical simulation/IK success, long-training equality or a full benchmark result. The handover includes separate current tiny training/resume and eval/session bitwise fixtures.
The bundled metadata initializes model structures on reload; separate backbone caches are unnecessary for inference. New training requires the full upstream pretrained DINOv3/SigLIP snapshots specified by GitHub manifests. No simulator, official assets, original training data or optimizer state is included.
Upstream DINOv3, SigLIP and RoboTwin terms continue to apply; this release does not add a blanket license.