ThinkWAM checkpoints

Project page · Code · Evaluation instructions

ThinkWAM inference checkpoints used for the paper's reported simulation results. Each .pt contains the model weights and its embedded configuration; optimizer, scheduler and RNG states are excluded. The original tensor values, storage precision and embedded configuration are preserved.

Benchmark Paper results Checkpoint
LIBERO / LIBERO-Plus Tables 1 and 2 libero_thinkwam_checkpoint.pt
RoboTwin 2.0 Table 3 robotwin2.0_thinkwam_checkpoint.pt
RoboCasa24 Table 4 robocasa24_thinkwam_checkpoint.pt

Training data and paper recipes

LIBERO training used the standard LIBERO demonstrations only. LIBERO-Plus was used for evaluation, with no LIBERO-Plus training or fine-tuning. RoboTwin training used Clean demonstrations only; Random was an evaluation condition. The serialized LIBERO plus_dataset_root and plus_mixture_fraction entries, and the RoboTwin random_dataset_root entry, are legacy configuration fields that were not used to supply training data for these checkpoints.

The authors also confirm that task-difficulty weighting, task curricula and RoboCasa event-based oversampling were not used for the paper runs. Those legacy sampling fields are absent from the active paper recipes and are excluded when the released loader constructs the inference configuration.

The authors confirm that the effective residual-loss coefficient lambda_res was 0.2 for all three paper training runs, as specified in the manuscript and appendix. The embedded video_residual_loss_weight entries below are stale configuration values, not the coefficients applied in those runs:

Benchmark Effective paper-run coefficient Original embedded value
LIBERO 0.2 0.8
RoboTwin 2.0 0.2 0.8
RoboCasa24 0.2 1.0

This clarification concerns training provenance; the checkpoint files and their original metadata are preserved. The residual-loss coefficient is not used in policy inference. The paper recipes already set it to 0.2.

The effective learning-rate schedules follow Appendix Table 9, as confirmed by the authors. The different learning-rate values serialized in the released artifacts are stale records and do not specify the applied paper-run schedules:

Training stage Peak learning rate Minimum learning rate
LIBERO P0 1e-4 1e-5
LIBERO P1 1e-4 1e-5
RoboTwin 2.0 5e-5 5e-6
RoboCasa24 P0 1e-4 1e-5
RoboCasa24 P1 1e-4 1e-6

The paper recipes use the stage's common learning rate for the active parameter groups, with the configured warmup and cosine decay.

The five configs/paper/*.yaml recipes in the code repository describe the settings in the manuscript and appendix. The artifacts retain their original serialized configuration and step counters. A stored step counter should not be interpreted as a complete stage history without its run records.

Loading and evaluation

Use the ThinkWAM evaluation commands with --checkpoint-profile released, following the release guide. The compatible loader checks the pinned artifact/config identity and strict-loads all policy parameters, excluding only LIBERO's eight inactive legacy phase-head parameters whose memory contribution is configured to zero. The two other checkpoints have no phase head.

The release evaluation path preserves the original serialized metadata and records paper_recipe_match: false. This flag describes the strict recipe-metadata check; it does not mean that the checkpoint was not used for the paper results listed above. Loading compatibility is not an independent reproduction of the paper's success rates. Full benchmark evaluations still require the simulators, original normalization statistics, assets and complete case schedules. See the code repository for validation coverage and limitations.

The V-JEPA encoder is embedded in the policy. Supply a local Qwen2.5-0.5B model and tokenizer directory with --qwen-path; runtime asset paths are relocatable. File sizes and SHA256 checksums are recorded in checkpoint_manifest.json and SHA256SUMS.

License

This checkpoint distribution is released under Apache License 2.0. Embedded Qwen2.5-0.5B tensors retain Apache 2.0 terms, and V-JEPA 2.1 tensors retain their MIT terms and attribution; see NOTICE. The separately released code and third-party code components retain their respective licenses.

Downloads last month

-

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