AllenAI has released ManiFlow-tb, a low-level action model designed for the ToyBlocks robot block-stacking project using simulated Franka arm data.
By Hermes-Vector AI Desk
AllenAI has released ManiFlow-tb, a low-level action model developed for ToyBlocks, a project focused on robot block-stacking. The model is designed for manipulating blocks on a table, specifically performing pick-and-place and in-place rotation of three blocks using simulated Franka arm data.
ManiFlow-tb utilizes the ManiFlowMultiModal architecture with a DiTX action head. According to the model documentation on Hugging Face, the model contains 322.5M parameters and is stored in bf16 format. The architecture includes 12 layers, 8 heads, and a width of 768.
Inputs for the model include:
The model outputs a chunk of 25 executed actions with a prediction horizon of 50. These are 9-dimensional absolute actions.
Training for the model involved 200,000 steps on 8 GPUs using three ToyBlocks-generated LeRobot datasets of tabletop block manipulation. The training utilized the AdamW optimizer and a cosine schedule with 500 warmup steps.
Evaluation was conducted through closed-loop simulation in ToyBlocks using a held-out validation set of three-block tabletop tasks. The evaluation measured the low-level action model's performance using an oracle high-level policy.
According to the technical documentation, success is defined as all goal predicates holding with the blocks settled and the gripper open. In four separate evaluation runs, the success rates were reported as follows:
| Run | Success Rate | Success at Episode End | Collisions |
|---|---|---|---|
| 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 documentation notes that the simulation and batched inference are not bitwise deterministic, meaning the same episode can result in different outcomes across different runs.
ManiFlow-tb provides a specialized model for low-level robotic manipulation tasks within the ToyBlocks framework. The model weights and the training code have been released under the Apache License 2.0.
Sources
This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.
The day's checked news, by email
One short email a day with the top stories. Free, and you can unsubscribe in one click.
Or follow for the latest news:
No comments yet. Start the discussion.
OpenAI has released GPT-6 Cyber, a model designed to find and validate exploitable vulnerabilities while remaining within researcher-set boundaries.
New approaches using mechanism design aim to incentivize AI models to follow rules and prevent deceptive behaviors like sandbagging.
