TILDE Teaches Robots From Foundation Models, Improving Results by Up to 4.1x

A team of researchers has presented TILDE (Teacher-In-the-Loop Distillation from EFMs), a way to transfer manipulation skills from large pretrained Embodied Foundation Models to small policies that have never undergone action pretraining. On three real-world robotics tasks, the method delivers a gain of up to 4.1x over standard behavior cloning. The authors also propose AutoDAgger, a pipeline that collects error-recovery data more than twice as fast as a human-in-the-loop setup.

Why action pretraining can't simply be repeated

EFMs that have gone through large-scale action pretraining generalize well to new situations and stay robust in closed loop: if something goes wrong, they correct the motion. That comes at a price. This kind of pretraining requires huge datasets collected from robots and serious compute, which most labs and companies simply don't have.

This raises a practical question: can you take the teacher's capabilities without retracing its path? The classic answer is behavior cloning, that is, copying demonstrations. It has a well-known weakness. The student trains on trajectories where everything goes right, but in the real world it makes mistakes and lands in states that never appeared in the demonstrations. Errors accumulate, and the policy falls apart.

Three sources of signal in one scheme

TILDE combines three kinds of learning: from demonstrations, from teacher supervision in the student's own states, and from the teacher's corrective behavior.

• Demonstrations set the baseline behavior.
• On-policy distillation provides teacher guidance where the student actually ends up on its own, rather than where the operator happened to be when recording the data.
• The teacher itself executes the recovery trajectories. When the student goes off track, the teacher takes control and shows how to get back to the task.

In essence, this extends the DAgger idea, in which an expert labels the states visited by the policy being trained. The difference is that the expert here is an EFM, not a human.

AutoDAgger: the operator leaves the loop

The bottleneck of standard DAgger is the human. Someone has to watch the robot, spot a failure in time, and take over control manually. AutoDAgger removes that role. A robotics reward model tracks task progress, and when progress stalls or reverses, it triggers a teacher intervention. The teacher executes a recovery trajectory, which goes into the training set with no observer and no manual control.

The result is data collection for these trajectories that is more than 2x faster than the human-in-the-loop method.

Key parameters

• Method: distillation from an EFM with the teacher in the loop (TILDE)
• Data collection: AutoDAgger with a robotics reward model as a progress monitor
• Speedup in collecting recovery trajectories: more than 2x relative to human-in-the-loop
• Quality gain: up to 4.1x versus behavior cloning
• Evaluation: three real-world tasks
• Additionally: support for cross-architecture distillation

What to do with this in practice

The most notable point is cross-architecture support. The student does not have to mirror the teacher's design, so a compact policy tailored to specific hardware and a latency budget can be trained from a large model without forcing the two architectures to match. For teams that have access to a strong EFM but lack the resources for their own pretraining, this is a realistic route to a working policy.

The second practical takeaway concerns the cost of data. If a human no longer has to stand watch at the robot, collecting corrective data turns from manual labor into a background process that can run in parallel on several setups.

The limitations are also clear. The whole scheme rests on the reward model: if it raises the alarm late or raises a false one, the teacher will step in at the wrong time. Data collection also requires a running teacher, and three tasks are still a small sample from which to draw conclusions about generality. Still, the direction looks convincing: large models are becoming not the end product but a trainer for small policies that are cheap to run.