Release and availability
The Allen Institute for AI (Ai2) has published the weights for Molmo2-4B-tb and ManiFlow-tb, two models developed for manipulating blocks on a tabletop. Both checkpoints are publicly accessible through their Hugging Face repositories under the Apache 2.0 license. The Molmo2-4B-tb model card gives November 2026 as the model date; the public files and model cards establish that the weights are available, rather than just announced.
The models are part of ToyBlocks, an open simulation framework built around IsaacSim and IsaacLab. ToyBlocks generates demonstration data at scale without requiring time-consuming teleoperation. Researchers can vary the kinds of demonstrations generated—for example, emphasizing pushing or flipping—and then evaluate policies on tasks or scenes excluded from training.
ToyBlocks also provides code for data generation, training and evaluation. Its README describes datasets recorded in LeRobot format and a workflow for testing policies in closed-loop simulation. The framework can evaluate policies on newly generated scenes or on the initial states of recorded episodes. This makes it possible to examine how a policy behaves under controlled conditions; it does not, by itself, establish how well the models will perform in other environments.
How the two-model system works
Molmo2-4B-tb is the high-level model. The model card describes it as a vision-language model with about 4.5 billion parameters. It takes a text goal and one 640 × 360 image from the robot’s front camera. It then reports progress toward the goal, names a next step and points to a block or location in the image. Its discrete actions include picking up or placing a block, flipping or rotating it, and indicating that the task is done.
ManiFlow-tb handles low-level action generation. Its model card lists 322.5 million parameters and inputs consisting of front and gripper camera images, a 9-dimensional robot state, a task name and a target heatmap from the high-level policy. It outputs a chunk of 25 executed actions from a prediction horizon of 50. The action representation includes a play-mode value, so the system that runs the policy must handle play, pause and skill-completion states.
The separation between planning and action generation helps distinguish two sources of failure: choosing an unsuitable next step and failing to execute a suitable step. It also means the reported success rate depends on the whole system configuration. A result with a motion planner as the low-level executor does not measure the same combination as a result with ManiFlow-tb.
The roles of Molmo2-4B-tb and ManiFlow-tb
| Feature | Molmo2-4B-tb | ManiFlow-tb |
|---|---|---|
| Role | High-level planner | Low-level action policy |
| Parameters | About 4.5 billion | 322.5 million |
| Inputs | Front-camera image and text goal | Front and gripper images, robot state, task name and target heatmap |
| Output | Plan, discrete action and target point | Chunk of 25 executed actions from a prediction horizon of 50 |
Reported evaluation results
Ai2 reports closed-loop success rates for the combined system in ToyBlocks simulation. Each of the Easy, Medium and Hard test sets contains 200 episodes. The tests vary in difficulty: Easy covers basic stacking, Medium includes reorienting blocks as well as stacking, and Hard adds searching for blocks. The model card also separates blocks seen during training from held-out blocks.
On Easy episodes with blocks seen during training, Molmo2-4B-tb paired with a motion planner succeeds in 97.5% of episodes. With ManiFlow-tb as the low-level executor, the reported rate is 55.5%. On held-out blocks, the corresponding Easy results are 81.0% and 44.5%. These are system-level results: changing the low-level executor changes the system being evaluated.
The reported rates fall on the more demanding tests. With ManiFlow-tb, success is 35.5% on Medium episodes with seen blocks and 17.0% on Hard episodes with seen blocks. On held-out blocks, the rates are 12.5% and 10.0%, respectively. These figures describe performance within the specified simulated tasks; they are not a general measure of manipulation ability.
Ai2 also reports 6 successful trials out of 10 for Molmo2-4B-tb paired with ManiFlow-tb on a real robot. The small number of trials makes this a limited demonstration, not a reliable estimate of performance across robots, objects or operating conditions. The result should not be read as a guarantee that the system will achieve the same rate in another setup.
Closed-loop success rates reported for ToyBlocks simulation
Success rate by test difficulty, training familiarity of the blocks, and low-level executor. Each test set contains 200 episodes.
| Difficulty | Motion planner, seen blocks | Motion planner, held-out blocks | ManiFlow, seen blocks | ManiFlow, held-out blocks |
|---|---|---|---|---|
| Easy | 97.5% | 81.0% | 55.5% | 44.5% |
| Medium | 53.5% | 36.5% | 35.5% | 12.5% |
| Hard | 31.5% | 14.5% | 17.0% | 10.0% |
What the ManiFlow evaluation isolates
Ai2 evaluates ManiFlow-tb separately in simulation using an oracle high-level plan drawn from the recorded task. This setup reduces the contribution of high-level planning errors and focuses the test on the low-level action policy. It uses 128 episodes from a held-out validation set of three-block tabletop tasks, each starting from the initial state of a recorded episode.
The model card reports four runs on the same episodes. Success ranges from 69 to 76 episodes out of 128, or about 54% to 59%. The evaluation uses an episode limit of 194 seconds, set to three times the longest expert demonstration in that evaluation. With the default 30-second limit, the card says most episodes time out before completion; that setting measures the time limit’s effect as well as the policy’s behavior.
The four runs do not produce identical results. Ai2 attributes the differences to non-bitwise-deterministic simulation and batched inference, and reports that agreement on individual episodes varies between runs. The range therefore describes variability in this particular evaluation, not a statistical confidence interval or an independently replicated result.
These oracle-plan results should not be confused with the full system’s performance. They test ManiFlow-tb under a supplied high-level plan; they do not show how the same policy performs when Molmo2-4B-tb must choose every next step.
ManiFlow-tb with an oracle high-level plan
Ai2 reports four closed-loop runs on the same 128 validation episodes, using a 194-second episode limit.
| Run | Successful episodes | Success rate | Episodes successful at 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 |
Technical requirements and limits of use
The Molmo2-4B-tb card estimates that inference requires about 20 GB of GPU memory in bfloat16. The published checkpoint is in Molmo2’s native format, rather than a Transformers-ready format, and the card directs users to ToyBlocks for inference and evaluation code. ManiFlow-tb is distributed with its model configuration and training-data statistics; evaluating it uses ToyBlocks’ policy server.
Neither checkpoint is a general-purpose robot controller. The model cards describe a narrow setting: moving blocks on a table, with the ManiFlow policy trained on simulated Franka-arm data. Molmo2-4B-tb predicts a plan, action and target point; it does not directly issue generic motor commands. Use beyond the documented setup has not been established by these evaluations.
The model cards list Apache 2.0 for the released weights, but the Molmo2-4B-tb card warns that its base model was trained on third-party datasets subject to academic and non-commercial research conditions. A permissive license on a checkpoint does not automatically remove conditions associated with its base model or data. Users should assess the applicable terms for their intended use.
The results also do not establish robust performance in unstructured settings. The real-robot demonstration covers 10 trials, while the simulation tests use a controlled tabletop environment and specified blocks and tasks. Neither result supports a claim of safety or reliability for other robots, objects or tasks.
Why ToyBlocks matters for robotics research
ToyBlocks is designed as a controlled environment for studying robot learning. Its data-generation pipeline can create demonstrations without requiring each episode to be teleoperated. Researchers can alter the distribution of experience—for example, generating more demonstrations of pushing or flipping—and then test whether a policy handles tasks kept out of training.
That setup can help researchers trace a trained policy’s outcomes back to the data-generation process and investigate where a system fails. ToyBlocks describes a workflow in which failed states can be sent back through data generation to plan expert demonstrations or recovery behavior. The framework provides tools for this investigation, but it does not mean that every generated demonstration will improve a policy or that performance will transfer automatically to the physical world.
The release is therefore most useful as a reproducible research setting for examining the interaction between high-level planning and low-level control. Its value lies in having a controllable simulation, data-generation tools and evaluation procedures alongside released checkpoints. The reported results remain specific to those models, tasks, configurations and test conditions.
ToyBlocks data and evaluation workflow
A typical research workflow described by ToyBlocks.
- 01Install ToyBlocks and initialize its required submodules.
- 02Download the simulation assets and configure the environment.
- 03Generate demonstration data with ToyBlocks’ data-generation pipeline.
- 04Train a low-level policy such as ManiFlow-tb on the generated data.
- 05Evaluate the policy in closed-loop simulation, or use the project’s real-robot setup where applicable.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction