In-context imitation learning allows robots to acquire skills from demonstrations, yet one-shot trajectory generation remains fragile under environmental variation. We propose SAIL, a framework that reframes robot imitation as an iterative refinement problem capable of scaling with test-time compute. SAIL utilizes Monte Carlo Tree Search, where each node is a complete trajectory and edges correspond to trajectory refinements. The process is guided by three core components: an automated archive of successful trajectories for contextually relevant retrieval, a vision language model-based scoring mechanism for trajectory evaluation, and a step-level feedback that provides trajectory-aligned scores for iterative refinement. Experiments across six diverse manipulation tasks in simulation and real-world validation clearly demonstrate that increasing test-time compute consistently improves success rates, achieving up to 95% on complex tasks. Our results suggest that trajectory-level test-time scaling is a robust path toward more generalizable robotic agents.
Figures & tables
Fig. 2: Method overview. SAIL refines trajectories at test time via MCTS, where each node is a complete trajectory proposed by a policy VLM for a seed. (1) A shared trajectory archive stores successful rollouts across seeds and retrieves visually similar trajectories as in-context demonstrations. (2) Each proposal is executed and scored by a scoring VLM that (i) decomposes the task into ordered subtasks from one demo and (ii) estimates per-frame completion to yield a scalar node value. (3) The scoring VLM aligns the progress scores to waypoints to provide step-level feedback for the next refinement.
Method
Nodes
Success Rate
HOB
HOP
BOR
DO
LC
MRL
Avg
Single rollout
1
0.40
0.40
0.40
0.10
0.15
0.05
0.25
Breadth-search
15
0.85
0.50
0.60
0.35
0.60
0.15
0.51
Depth-search
15
0.45
0.50
0.80
0.05
0.30
0.10
0.37
Ours
6
0.80
0.55
1.00
0.20
0.50
0.25
0.55
15
0.90
0.70
1.00
0.40
0.50
0.40
0.65
TABLE I: Success rates over 20 seeds under different test-time compute budgets measured by expanded MCTS nodes. Gray rows indicate the 15-node budget for fair comparison. Bold denotes the best result among the 15-node methods, and underline denotes the best overall.
Setting
Success Rate
HOB
HOP
BOR
DO
LC
MRL
Avg
Ours ( K=1 )
0.90
0.70
1.00
0.40
0.50
0.40
0.65
Retrieval Strategy ( K=1 )
Fixed demonstration
0.75
0.40
0.85
0.05
0.50
0.15
0.45
Random retrieval
0.75
0.50
0.90
0.15
0.55
0.15
0.50
Retrieval Strategy ( K=3 )
TABLE II: Ablation study with a fixed MCTS budget of 15 nodes (success rate). We compare retrieval strategies and feedback modalities for refinement.
Fig. 3: Digital-twin experimental setup for the BlockIntoBowl task. The task consists of three sequential steps: (1) grasp the blue block, (2) move to the bowl, and (3) release the block into the bowl. The left panel shows the simulation environment, and the right panel shows the corresponding real-world setup.
Method
Success Rate
MCTS-based Trajectory Refinement
5 / 6
Sim2Real Policy Distillation
5 / 6
TABLE III: Real-world BlockIntoBowl success rates over 6 trials.
Global College, Shanghai Jiao Tong University, Shanghai, China · Shanghai Institute of Physical Intelligence and Robotics (PAIR), Shanghai, China · Duke Kunshan University, Jiangsu, China
University of Electronic Science and Technology of China · Huazhong University of Science and Technology · Southern University of Science and Technology +2