ManiFlow-tb

ManiFlow-tb is the low-level action model for ToyBlocks, a robot block-stacking project. It is trained on simulated Franka arm data for manipulating blocks on a table (pick-and-place and in-place rotation of three blocks). It is trained with the ManiFlow training code at allenai/maniflow-tb and evaluated through ToyBlocks' policy server.

Quick links:

• 📝 ToyBlocks blog post
• 💻 ToyBlocks code
• 🛠️ ManiFlow training code
• 📊 ToyBlocks data

Model

• Architecture: ManiFlowMultiModal with a DiTX action head (12 layers, 8 heads, width 768), 322.5M parameters, stored in bf16 (safetensors).
• Inputs: front and gripper RGB images (224 x 224, CLIP ViT-B/16 encoder), a 9-dim robot state, the task name (T5-small, language_mode: action_only) and a target heatmap token from the high-level policy (target_injection_mode: token, encoded by the shared RGB encoder, heatmap_encoder_type: shared_rgb).
• Outputs: a chunk of 25 executed actions (prediction horizon 50) of 9-dim absolute actions (action_mode: abs). The last dimension is the play mode (1 play, 0 pause, -1 skill done), because the model was trained with pause_on_attempt: true. The play mode needs to be fed back to the model when 1 or 0.
• Flow matching with consistency training, 12 integration steps at inference.

Training

• Data: three ToyBlocks-generated LeRobot datasets of tabletop block manipulation.
• 200,000 steps on 8 GPUs, per-device batch 128, AdamW (lr 1e-4, betas 0.9/0.95, weight decay 1e-3), cosine schedule with 500 warmup steps, bf16 mixed precision. Full settings are in config.yaml.
• The weights are the online (non-EMA) weights of the final step (200,000).

Files

The repository mirrors a training run directory, which is the layout ToyBlocks expects:

file	content
checkpoints/steps_200000_model.safetensors	model weights
config.yaml	training config (ToyBlocks reads the model and action settings from it)
stats.json	normalization statistics of the training data

Usage

Download the repository:

hf download allenai/maniflow-tb --local-dir maniflow-tb

Evaluate it with ToyBlocks (the server finds config.yaml next to the checkpoints/ directory):

python scripts/launch_eval.py --policy maniflow \
  --model-path maniflow-tb/checkpoints/steps_200000_model.safetensors \
  --dataset-path <ToyBlocks eval dataset> --expert-timeout-scale 3.0 ...

Training code: allenai/maniflow-tb (scripts/train_maniflow.py). ToyBlocks also uses this code as its baselines/hvla package to load the model.

Evaluation

Closed-loop evaluation in ToyBlocks simulation with its own eval pipeline (scripts/launch_eval.py --policy maniflow):

• 128 episodes from a held-out validation set of three-block tabletop tasks, each started from the initial state of a recorded episode.
• Oracle high-level policy: the high-level plan comes from the recorded task, so the numbers measure this low-level action model alone.
• Success: all goal predicates hold with the blocks settled and the gripper open. An episode ends when the high-level plan is complete, on a collision, or at the time limit.
• Time limit: 3x the longest expert demonstration (--expert-timeout-scale 3.0, 194 s here). With the default 30 s, most episodes time out before the task is finished (26 to 28 of 128 succeed), which measures the limit rather than the model.

run	success	success at episode end	collision
1	73/128 (57%)	64	32
2	75/128 (59%)	67	32
3	76/128 (59%)	66	30
4	69/128 (54%)	63	38

The four runs use the same episodes and settings; runs 1, 2 and 4 spread the episodes over 7 eval GPUs, run 3 over 1. They differ because the simulation and batched inference are not bitwise deterministic. The same episode can flip between success and failure from run to run: two runs agree on 82 to 97 of the 128 episodes.

License

The model weights and the training code are released under the Apache License 2.0.