--- pipeline_tag: robotics tags: - imitation-learning - flow-matching - diffusion-policy - so-101 - robot-imitation-glue --- # dit_block_2 — DiT flow-matching policy, cup-into-ring Single-camera Diffusion-Transformer flow policy for the SO-10x task *pick up the cup, place it in the ring*. Trained on the 2026-11-15 clean wrist retake of the cup dataset. **This is not a `from_pretrained` checkpoint.** It is one raw `torch.save` dict written by `robot_imitation_glue.baseline.train_dit`. lerobot's policy registry cannot open it — there is no `config.json` / `model.safetensors` here. Load it through `DiTAgent` (below). ## File | file | bytes | md5 | |---|---|---| | `policy.ckpt` | 66,705,733 | `669479fdbcfa196c7c251c4bf9e7c549` | Contents: `config`, `image_size`, `crop_ratio`, `model_state_dict` (16,660,576 params), `proprio_normalizer_state_dict`, `action_normalizer_state_dict`. The two normalisers ship with the weights — the model is unusable without them, and nothing else (no dataset paths, no credentials) is inside the file. ## Task interface - **State / action**: 6-D absolute Feetech positions, dataset order `shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper`. Dims 0-4 are **degrees**; dim 5 is **% open** (the gripper servo is hard-wired to `RANGE_0_100`). Not one unit — do not scale the vector. - **Vision**: one camera, `observation.images.front`, recorded 480×640 AV1 → resized to 128×128 with a 0.9 centre crop. - **Chunk**: `prediction_horizon = 48` counts **frames**, and the dataset is **30 fps**, so one chunk is 1.6 s of motion. A control loop must therefore run at 30 fps; at 10 fps the same chunk stretches to 4.8 s. ## Architecture | | | |---|---| | hidden dim | 256 | | layers / heads | 4 / 4 | | camera encoder | resnet18, ImageNet-pretrained, frozen statistics shipped as `encoders.0.mean/std` | | prediction | flow matching, Euler sampler (10 steps is what the deployment script uses) | | params | 16,660,576 | ## Training | | | |---|---| | dataset | [`anschrur/record-test_2_20261115_115045`](https://huggingface.co/datasets/anschrur/record-test_2_20261115_115045) — 30 episodes, 13,218 frames, 30 fps, 1 task | | optimiser steps | 41,200 = 200 epochs × floor(13,218 / 64), `drop_last=True` | | batch / optimiser | 64 / AdamW, lr 1e-4, weight decay 1e-4 | | schedule | 500-step linear warmup → cosine to 1 % of base (final lr 1e-6), grad-clip 10, seed 42 | | final train loss | 0.018505 (epoch 200/200) | | hardware / wall clock | 1× RTX 5090 (imec iLab.t GPULab cluster 9), 1.7 h | | written | 2026-11-15 13:28, `/project_ghent/dit/outputs/dit_cup_2/policy.ckpt` | `--save-every 20` rewrites the *same* `policy.ckpt` path, so these are the epoch-200 weights only — no intermediate snapshots and no epoch sweep to plot. The dataset declares no held-out split, so `train_loss` is a training-set number, not an evaluation result. ## What changed against `anschrur/dit-cup` Identical architecture and hyperparameters, with the epoch count raised 30 → 200 so the **step** totals match (40,710 then vs 41,200 now); the dataset is the only variable. `dit-cup` trained on `record-test_2`, whose `wrist_roll` labels carry a ~170° offset in 32 of 200 episodes. The retake removed it, and the difference shows up in the action normaliser stored inside the file (degrees for dims 0-4, % open for dim 5): | dim | `dit-cup` mean / std | `dit_block_2` mean / std | |---|---|---| | shoulder_pan | −2.94 / 18.42 | 0.79 / 16.15 | | shoulder_lift | −20.40 / 46.84 | 79.98 / 51.82 | | elbow_flex | 28.94 / 36.92 | −71.29 / 43.93 | | wrist_flex | 60.48 / 14.90 | 73.85 / 13.39 | | **wrist_roll** | **−144.60 / 69.53** | **4.37 / 20.51** | | gripper | 8.01 / 13.20 | 6.85 / 11.46 | `wrist_roll` std falls 69.5 → 20.5 once the offset is gone. The other dims move too because the retake is a different 30-episode recording, not a relabelled subset of the old one — those four joints are not directly comparable. ## Loading ```python from robot_imitation_glue.baseline.dit_agent import DiTAgent # plain clone on sys.path is enough agent = DiTAgent( checkpoint_path="policy.ckpt", camera_key_mapping={"front": "observation.images.front"}, # live camera key -> policy key state_key="observation.state", device="cpu", n_sampling_steps=10, ) ``` Closed-loop on an arm — the repo is public, so this needs no token: ```bash python dit_on_robot.py --repo-id anschrur/dit_block_2 --fps 30 --dry-run # arm must NOT move on this one ``` `dit_on_robot.py` (thesis glue, not published here) wires `DiTAgent` to lerobot's `SOFollower` with a wrap-aware per-step delta guard, because lerobot's own `max_relative_target` clamp mishandles the ±180° `wrist_roll` boundary. ## Not measured No open-loop action MSE, no evaluation episodes, no success rate, and no head-to-head against `dit-cup` beyond the normaliser stats above. The wrist_roll label defect in the previous dataset is the reason this run exists; whether it improves closed-loop behaviour is what the next robot rollout has to answer.