SLA is a robot policy trained to connect spatial cues, language, and actions through continued pre-training. Across simulation and real-world tests, it is evaluated on visual changes, kitchen work, bimanual tasks, and three robot embodiments, offering a practical test of whether one learned system can transfer beyond a single setup.
What did the researchers build?
The researchers built SLA, short for Spatial-Language-Action, as a robot policy intended to use visual information, instructions, and movement in one system. Rather than starting from scratch for each benchmark, they initialized it from a pretrained policy and continued pre-training before adapting it to specific tasks. That adaptation, called post-training, uses task demonstrations to tune a model for a particular robot or evaluation.
The evaluation spans both simulated and physical settings. In simulation, SLA is tested on visual and language perturbations, bimanual manipulation, and kitchen tasks. In the physical-robot benchmark, a single jointly post-trained model is tested across three embodiments: ARX X5, Piper, and Piper X. The authors also describe real-world deployment tasks designed to test coordination, generalization, adaptation to different robot bodies, and instruction following.
The paper’s approach addresses a practical tension: a general robot policy needs broad skills, but a robot deployed in a specific operation still needs targeted adaptation. SLA’s training setup combines both stages—continued pre-training for the general model, followed by benchmark- or deployment-specific post-training.
What results did SLA achieve?
The available results report strong performance on several individual RoboCasa-Panda tasks. SLA reached a 97.3% success rate on CloseDrawer, 95.3% on OpenDoubleDoor, and 93.3% on TurnOnMicrowave. These examples cover articulated-object interaction and appliance operation, rather than only simple object pickup. RoboCasa-Panda includes 24 kitchen tasks, with each task tested over three evaluation seeds and 50 trials per seed.
The other benchmarks are designed to measure different aspects of capability. LIBERO-Plus evaluates robustness to seven changes, including camera, robot, language, lighting, and object layout. RoboTwin 2.0 compares performance under clean conditions with performance when the scene is randomized. RoboDojo-Real measures both partial task progress on a 0–100 scale and complete task success across three physical robot types.
The provided results do not include overall success rates for these benchmarks or numerical comparisons against competing policies. The task-level RoboCasa figures therefore show where SLA can perform well, but they are not enough to establish an overall ranking or prove that continual pre-training beats other training approaches.
How does SLA work?
SLA’s training has two main stages. First, the team continues pre-training from an existing pretrained policy. This stage is intended to build on an already-trained model rather than teach every capability from a blank start. Second, the model is post-trained for a benchmark or deployment using demonstrations for its target tasks. The benchmark settings differ in the number of training steps and batch sizes, so they are not identical training runs.
For some real-world episodes, the policy also receives a short visual history. The memory contains the initial scene frame plus one frame per second from the preceding eight seconds, ordered from oldest to newest. The initial frame remains available throughout the episode to preserve the original scene layout; the later frames provide evidence of recent motion and task progress. During the first eight seconds, any unavailable history frames are filled with copies of the initial frame. Wrist cameras provide current images, while the saved history comes from the main camera.
The evaluation scale is substantial in some benchmarks, though trial counts measure how extensively a model was tested—not how often it succeeded. The table summarizes the stated protocols.
| Benchmark | Training and evaluation setup | Stated evaluation scale |
|---|---|---|
| LIBERO-Plus | Standard LIBERO demonstrations; tests seven perturbation dimensions | 10,030 rollouts |
| RoboTwin 2.0 | 50 tasks; 50 clean demonstrations per task; clean and randomized settings | 100 trials per task in each setting |
| RoboCasa-Panda | 24 tasks; 50 demonstrations per task; three evaluation seeds | 3,600 trials |
| RoboDojo-Real | 18 tasks across ARX X5, Piper, and Piper X | Six tasks per embodiment |
Post-training runs range from 30,000 steps for LIBERO-Plus and RoboTwin 2.0 to 60,000 for RoboCasa-Panda. RoboDojo-Real uses 40,000 steps, while the real-world deployment experiments use 20,000. These details help describe the effort behind each evaluation, but they do not by themselves show how much data or compute the continual pre-training stage required.
Why does SLA matter for robotics?
Many robot deployments are narrow: a model is trained for one workcell, one camera arrangement, or one robot arm, then struggles when the surroundings change. SLA’s evaluation plan targets that gap directly. It tests visual and language perturbations in simulation, randomizes scenes, includes two-handed tasks, and measures a single model across different physical robot embodiments. For robot operators, those are relevant stress tests because deployment conditions rarely remain perfectly fixed.
Cross-embodiment training is especially useful if one policy can be adapted to different hardware without rebuilding the full system each time. The reported RoboDojo-Real setup tests that possibility across three robots, although the supplied results do not give enough aggregate scores to quantify the transfer benefit. Memory of recent camera frames is also a practical design choice: manipulation often depends on knowing what moved, not just what the scene looked like at the start.
SLA is not a turnkey product, and the benchmark results do not settle procurement questions such as cycle time, reliability over long shifts, or integration cost. Still, buyers comparing automation options can use this work as a useful reference point when considering used robots for sale and the software needed to make different platforms perform flexible tasks.
What are the limitations and open questions?
The reported task-level successes are encouraging, but they do not show SLA’s overall performance across all four benchmarks. The available results also do not provide baseline comparisons, so the gains over the starting pretrained policy or other approaches cannot be assessed here. That makes it difficult to tell whether continued pre-training, post-training data, model design, or some combination explains the reported outcomes.
The physical-robot evaluation covers three embodiments and selected real-world tasks, not every robot configuration or operating environment. More evidence is needed on repeatability, recovery from failures, speed, and performance over extended operation. The paper describes the memory-frame strategy and benchmark training schedules, but the supplied material does not specify the continual pre-training corpus, its scale, or the compute cost. Those details matter when judging whether the method can be reproduced and economically deployed.
What do readers ask about SLA?
What does SLA stand for? It stands for Spatial-Language-Action, describing the combination of spatial information, language instructions, and robot actions.
Does SLA work on real robots? The evaluation includes physical deployments and a real-world benchmark spanning ARX X5, Piper, and Piper X robots.
How does SLA handle changing scenes? Its tests include camera and layout changes, randomized environments, and a short visual history that helps track recent task progress.
Did SLA outperform other robot policies? The available results include selected task success rates but not comparative baseline scores, so they do not establish an overall win.
What is the practical takeaway?
SLA combines continued pre-training with task-specific adaptation and evaluates the resulting policy across diverse simulated and physical manipulation settings. Its selected kitchen-task results are strong, while broader comparisons and deployment details are needed to judge how much it improves real-world robot performance.