Glossary
What Is Post-Training in Robot Learning?
Last reviewed: September 2026
Short answer
Post-training (also called fine-tuning) is the final stage of a three-stage robot foundation model training recipe: after broad pretraining and, in some recipes, an aligned mid-training stage, post-training narrows a model down to one specific deployment task, robot, or environment using a written skill spec. Where pretraining optimizes for breadth and scale, post-training optimizes for reliable performance on the exact task the robot will run in production.
01
Why it matters
A pretrained (and, where used, mid-trained) representation is broad but not tuned to any single deployment. Post-training is where that general representation gets narrowed to the task, robot, and environment it actually has to perform in — the stage that determines whether a model succeeds on the floor, not just on a benchmark. Because post-training operates on a written skill spec for one target deployment rather than broad coverage, it typically uses far less data than pretraining, but that data has to be captured on or matched closely to the target task and edge cases, or the earlier stages’ breadth does not translate into reliable execution.
02
How Firsthand handles it
Firsthand captures the egocentric video, depth, 3D hand pose, and action-segment data used for post-training — collected fast, on demand, and to a written spec, then delivered in RLDS, LeRobot format, or WebDataset with a buyer-owned commercial license. Because post-training data is scoped to one deployment task rather than broad coverage, the spec that matters most is the target task’s edge cases — variations in objects, layouts, lighting, and failure modes — rather than a volume target, and a custom collection can be scoped tightly around exactly the task a model is being specialized for.
03
Pretraining vs mid-training vs post-training, at a glance
| Stage | Goal | Typical data |
|---|---|---|
| Pretraining | Broad coverage and scale across tasks, scenes, and (often) embodiments | Large egocentric human-video or robot-demonstration corpora, e.g. EgoScale’s 20,854-hour pretraining set |
| Mid-training | Align human and robot embodiments on matched tasks and viewpoints | A smaller, tightly aligned human-plus-robot slice, e.g. EgoScale’s ≈50h human + ≈4h robot across 344 tabletop tasks |
| Post-training (fine-tuning) | Specialize the model to a specific deployment task, robot, or environment | A written skill spec collected against the target task and edge cases |
05
What a post-training data spec typically covers
| Spec element | Why it matters |
|---|---|
| Target task definition | Post-training narrows to one deployment task, so the task and its success criteria have to be written down precisely |
| Edge-case and failure-mode coverage | Reliability on the deployment task depends on capturing variation and recoverable failures, not just the nominal happy path |
| Target robot embodiment | Post-training data is captured on or matched to the robot that will actually run the task |
| Environment and object variation | The deployment environment’s layouts, lighting, and objects need direct coverage, since post-training does not rely on pretraining-scale breadth to fill gaps |
06
Frequently asked questions
- What is post-training in robot learning?
- Post-training (also called fine-tuning) is the final stage of a robot foundation model training recipe, where a model that has already been pretrained — and, in some recipes, mid-trained — is specialized to one specific deployment task, robot, or environment using a written skill spec.
- How does post-training differ from pretraining and mid-training?
- Pretraining maximizes breadth and scale across tasks, scenes, and often embodiments. Mid-training is a smaller, tightly aligned stage that adapts a pretrained representation to one target robot’s embodiment. Post-training narrows further still, specializing the model to a specific deployment task using a written skill spec. See pretraining vs mid-training vs post-training data for the full breakdown.
- What data is used for post-training?
- Data captured against a written skill spec for the exact deployment task, robot, and environment — including edge cases and failure modes — rather than the broad, large-scale corpora used for pretraining.
- How much post-training data do published recipes use?
- This site does not state a published hour count specifically for a post-training/fine-tuning stage. What is published is the earlier stages: EgoScale’s pretraining corpus is 20,854 hours, and its mid-training stage pairs roughly 50 hours of human video with about 4 hours of robot data across 344 tabletop tasks. Post-training data volume is driven by the target task’s edge-case coverage rather than a fixed hour target.
- Why does post-training need robot-embodiment data?
- Because post-training specializes a model to a specific deployment, the data has to reflect the robot that will actually run the task — its embodiment, viewpoint, and reach — rather than the broader, often embodiment-agnostic data used for pretraining.
- Is post-training the same as fine-tuning?
- Yes — post-training and fine-tuning refer to the same final stage: specializing a pretrained (and, where used, mid-trained) model to a specific deployment task, robot, or environment.
- Do all robot foundation models use a mid-training stage before post-training?
- Not necessarily. Some training recipes go directly from pretraining to post-training (fine-tuning). EgoScale is the published example of a recipe that explicitly inserts an aligned mid-training stage between the two. See what is mid-training for how that stage fits in.
- How does custom collection support post-training?
- A custom collection can be scoped tightly around one deployment task’s written spec — its objects, layouts, lighting, and edge cases — collected fast and on demand, rather than drawing from a broad pretraining corpus that was never built around that specific task.