ZUO66/imagewam-r1lite-insert-flower

ImageWAM: R1Lite insert flower

Final 20-epoch / step 11,440 inference checkpoint for the R1Lite right arm. This is a full fine-tune, not a LoRA adapter. It contains visual/action MoT weights and the trained proprio encoder, not optimizer or resumable state. No robot rollout success rate is claimed by this release.

Training and inference code and real-robot guide.

Files

• step_011440.pt: final visual/action/proprio inference weights.
• config.yaml: resolved training architecture/protocol with portable paths.
• dataset_stats.json: training normalization statistics; use these unchanged.
• manifest.json: byte counts and SHA256 checksums.

Protocol

98 episodes, 45,684 frames at 15 FPS, fixed dataset revision 5c4b884303cf7be101b0b8611ad6362b5441db4a. Head RGB and right-wrist RGB are independently resized to 224 x 224 and joined horizontally into 224 x 448. Actions are 7D absolute right-arm targets: six joints followed by right gripper. Current 7D proprio conditions the model. Past 16 actions predict future 16; q01/q99 normalization, boundary padding retained.

I2I source is (current latent + Gaussian noise) / sqrt(2). A2A source is normalized action history plus 0.5 * Gaussian noise on all seven dimensions. Visual and action endpoint-consistency objectives are enabled. Action-loss gradients into the visual conditioning branch are blocked. Qwen context length is 128; the current proprio adds one condition token. Training used 8 x A100 80 GB, microbatch 10, global batch 80, no accumulation, bf16, ZeRO-1, cosine LR 1e-4 to 1e-6, 572 warmup steps and 20 epochs. This run starts from base initialization, not from the trained cups checkpoint.

Download and inference

hf download ZUO66/imagewam-r1lite-insert-flower --local-dir checkpoints/r1lite_flower

Use scripts/r1lite/infer.py in HumanWAM with the matching model config, external FLUX.2 base weights/autoencoder and Qwen3-4B. The release does not redistribute these frozen dependencies. The full checkpoint overwrites the visual/action initialization; the action initializer is not needed for inference. The training instruction is: Use the right arm to pick up the flower from the transparent white vase on the left and insert it into the purple vase on the right. The full Qwen prompt adds the prefix: A video recorded from a robot's point of view executing the following instruction: . The inference example reads the instruction override and checkpoint filename from this release's config and manifest; do not use the cups instruction. Use one visual step and one action step with fresh noise for each policy call. Use executed command history rather than predicted-but-unexecuted actions.

Limitations and attribution

This checkpoint is task-, camera- and robot-specific. Validate offline first; do not send unverified predictions directly to a physical robot. The visual/action initialization derives from FLUX.2-klein-base-4B (Apache-2.0); ImageWAM implementation retains its upstream MIT license. Dataset attribution and terms remain with at237299966/r1lite_insert_flower.