Organizations: ByteDance Seed · Princeton University · Yale University · Carnegie Mellon University · Stanford University · University of Washington · University of California, Los Angeles
We study coding agents for long-horizon, dexterous robotics and ask whether their solutions can provide scalable supervision for learning general robot policies. To test this, we develop EMBODIEDSWE-BENCH, a simulation benchmark for coding agents spanning contact-rich manipulation, deformable objects, and long-horizon tasks requiring up to half an hour of continuous interaction. We find that frontier coding agents can solve complex long-horizon tasks and transfer prior solutions across both tasks and embodiments. We also design supporting tools that help agents more effectively solve these tasks. However, the resulting solutions require substantial iterative interaction and are typically specialized to individual task instances. We therefore introduce EMBODIEDSWE-GEN, which expands a single solution from coding agent into large diverse trajectories for training a VLA. VLA performance improves with more generated demonstrations, and agent-aided diversification improves generalization to held-out task variations. We also show that a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on real robot. Together, our framework uses coding agents to solve complex robotics tasks and turn verified solutions into scalable supervision for robot policies.
Figures & tables
Figure 1: Overview of our framework: Top: EmbodiedSWE-Bench contains diverse long-horizon robotics tasks. Bottom: coding agents solve these tasks and turn verified solutions into scalable supervision for general policies.
Figure 2: Representative successful trajectories from coding-agent solutions on EmbodiedSWE-Bench , spanning long-horizon manipulation across tasks and embodiments.
Figure 3: Base results on EmbodiedSWE-Bench . Left, middle: mean best-so-far score over 28 tasks against wall-clock and tokens; markers with vertical bars at hourly intervals / fixed token counts show mean ± SEM across tasks (one run per task). Right: mean final score with SEM across tasks; percentages give the fraction of tasks solved.
Figure 4: Score against cost.
Figure 5: Programmatic policies versus RL. Despite stronger supervision and extensive tuning, RL substantially underperforms coding-agent policies.
Figure 6: Evaluations on cross-task and cross-embodiment transfer (left two) and with additional tools (right two). Markers with vertical bars at hourly intervals show mean ± SEM across runs (2 models × 3 seeds per task for the generalization runs; one run per task for the tool runs).
Figure 7: Hierarchical diversification pipeline of EmbodiedSWE-Gen . Starting from one verified coding-agent solution, we diversify the scene , strategy , phase , dynamics , and visual to expand it into a broad set of training trajectories. Higher-level variations modify task structure, solution behavior, and intermediate-state coverage, while lower-level variations introduce physical, action, and visual perturbations to improve robustness and diversify execution.
Figure 8: One verified solution can be expanded into many diverse trajectories through hierarchical diversification, with only modest additional agent cost per variation and rapidly decreasing amortized cost at scale.
Figure 9: VLA finetuning with generated in-distribution demonstrations. Mean task scores and success rates at five training-set sizes. Overall performance improves as more generated demonstrations are provided.
Task
Script-only
Agent-aided (ours)
pen_holder
0.083±0.040
0.183±0.021
tool_packing
0.262±0.028
0.331±0.026
slice
0
0.220±0.017
bulb
0.026±0.013
0.236±0.081
nut_thread
0.025±0.006
0.300±0.027
clear_organic_objects
0.002±0.001
0.129±0.023
Table 1: Generalization to held-out variations. Both methods use the same number of training demonstrations.
Figure 10: Successful real-robot lamp disassembly rollout from a policy finetuned only on coding-agent-generated simulation data
Policy
Grasp shade
Place and release shade
Grasp bulb
Unscrew and Extract bulb
Sim-trained
100%
80%
30%
20%
Pretrained π0.5 -DROID
0%
0%
0%
0%
Table 2: Real-robot stage success rates over 10 trials. The finetuned policy achieves nontrivial success on the real robot, including completion of the challenging bulb-unscrewing stage, despite being trained exclusively on simulated demonstrations.
Appendix figures & tables25 assets
Supplementary material from the paper’s appendix.
Appendix
Suite
Task
Supported embodiments
Difficulty
Assembly
allen_bolt . Drive an Allen bolt into a threaded plate using an L-shaped key.
Franka; UFACTORY xArm7; Kinova Gen3
Medium
Assembly
bulb . Pick up a light bulb and screw it into its socket until fully seated.
Franka; UFACTORY xArm7; Kinova Gen3
Medium
Assembly
ikea_table . Attach four threaded legs to the corner studs of a tabletop.
Franka (bimanual)
Hard
Assembly
SO_101 . Assemble a robot arm.
Franka (bimanual)
Hard
Assembly
nut_thread . Pick up an M16 nut and thread it onto a fixed bolt.
Franka; UFACTORY xArm7; Kinova Gen3
Easy
Assembly
pc_gpu . Align a graphics card with a PCIe slot and press it into place.
Franka; UFACTORY xArm7; Kinova Gen3
Easy
Appendix
Table 3: EmbodiedSWE-Bench task catalog.
Type
Embodiment
Arm DOFs
End-effector
Notes
Franka
7
two-finger parallel gripper
fixed base; tabletop manipulation
Single arm
UFACTORY xArm7
7
linkage gripper or Panda hand
∼ 0.7 m reach
Kinova Gen3
7
Panda hand at the flange
∼ 0.9 m reach
Bimanual
Bimanual Franka
14
two parallel grippers
two Frankas with per-scene base placement; actions concatenated
Humanoid
Unitree G1
17
two three-finger hands (14 DOFs)
29-DOF humanoid; pelvis fixed in tabletop tasks; in locomanipulation tasks, the 12 leg DOFs are driven by a pretrained locomotion policy taking a base velocity command
Appendix
Table 4: Supported embodiments.
Controller
Commands
Action (per arm)
Used in
Operational-space control
torques (actuators at zero intrinsic stiffness)
six end-effector pose deltas
arm evaluations (default)
Differential IK
joint position targets tracked by the arm PD
six end-effector pose deltas
arm evaluations
Pink IK
joint position targets tracked by the arm PD
absolute end-effector pose
humanoid evaluations
Task-space impedance
torques (actuators at zero intrinsic stiffness)
six end-effector pose deltas
—
Direct joint targets
joint position targets
one target per joint
—
Appendix
Table 5: Provided controllers.
Figure 11: Best-so-far score versus wall-clock time, by difficulty tier.
Figure 12: Best-so-far score versus generated tokens, by difficulty tier.
Figure 13: Best-so-far score versus wall-clock time, by suite.
Figure 14: Best-so-far score versus generated tokens, by suite.
Figure 15: Best-so-far score versus wall-clock time, by embodiment.
Figure 16: Best-so-far score versus generated tokens, by embodiment.
Task
Stage terms (weight)
bulb
reach (0.1): pads at the bulb’s glass belly; lift (0.1): rise above the start height, full at 8 cm; transport (0.2): kernel on the bulb’s lateral offset from the nearest socket axis; upright (0.1): cosine between the bulb and socket axes; approach (0.2): when aligned, kernel on the height above the bore mouth; thread (0.3): when engaged, depth from the bore mouth to the seat.
nut_thread
reach (0.1): pads at the nut; lift (0.1): rise above the table, full at 6 cm; transport (0.2): lateral offset from the nearest bolt axis; upright (0.1): cosine between the nut and bolt axes; approach (0.2): when aligned, height above the thread start; thread (0.3): when engaged, depth from the thread start to the seat.
slice
reach (0.1): pads at the knife handle; taken (0.1): knife rise off its rest, full at 6 cm; carry (0.15): blade-edge centre over the food; align (0.15): blade normal parallel to the cut axis; press (0.2): the edge sample nearest the next uncut plane lies in that plane and is pressed toward the board; cut (0.3): fraction of planes released.
pen_holder
reach (0.1): pads at the nearest not-yet-inserted pen; lift (0.1): that pen’s rise off the table, full at 10 cm; carry (0.15): the pen’s bottom end over the holder axis; tip_up (0.1): pen axis along the holder axis, tip up; insert (0.25): when over the axis, pen bottom lowered below the rim toward the target depth; all_in (0.3): fraction of pens counted as inserted.
tool_packing
reach (0.1): pads at the nearest unstowed tool; lift (0.1): its rise off the surface, full at 8 cm; carry (0.2): tool over its assigned drawer; stowed (0.3): fraction of tools inside their drawers; drawers (0.2): mean drawer shut-ness; doors (0.1): mean door shut-ness.
Appendix
Table 15: Agent-designed dense rewards. Final stage terms and weights proposed by the coding agent for each task. Every term lies in [0,1] ; the potential is the weighted sum normalized by the total weight. Reach terms use a two-scale kernel 21[(1−tanh(d/0.30))+(1−tanh(d/0.05))] on the distance d between the finger pads and the grasp point; ramps are clipped to [0,1] .
Task
Stage terms
Training weights
bulb
reach (grasp pose at the neck waist), grasp, lift, transport, upright, approach, thread
reach 0.2, grasp 0.2, lift 0.6
nut_thread
reach (grasp pose on the upper half of the nut), grasp, lift, transport, upright, approach, thread
reach 0.2, grasp 0.2, lift 0.6; resumed with reach 0.1, grasp 0.1, lift 0.2, transport 0.3, upright 0.1, approach 0.2
slice
reach (grasp pose at the handle centre, fingers across the plate), grasp, taken, carry, align, press, cut
Table 16: Expert-tuned rewards. Stage terms defined by the expert for each task and the weights used in the final training runs. Because no policy learned the pick under a flat weighting (on pen_holder , grasp peaked at 0.04 and lift at 0.0003 over 300 iterations with seven terms at roughly 0.1 each), the weight was concentrated on the pick terms for the tasks whose first rubric stage is a lift; the remaining terms are defined and logged but weighted zero. tool_packing uses a per-item ladder in which the pick-and-place terms are gated on the assigned drawer being open.
Task
Additions
bulb
Grasp point moved from the glass belly to the neck waist (2.5 cm above the cap, 2 cm wide); pad offset re-measured; warm-start hover target 10 cm above the neck.
nut_thread
Pinch the upper half of the 13 mm nut (24 mm across flats); pad offset re-measured to the fingertips; warm-start hover target 10 cm above the nut; final run initialized from the bulb policy.
slice
Joint-delta actions with scale 0.15 rad per step (the arm PD caps joint speed at scale ⋅kp/kd , and 0.05 was too slow to reach the knife); lift target raised to 8 cm because the knife settles 2.1 cm into its notches after reset while the grader keeps the unsettled height; warm-start hover target 10 cm above the handle, with a bridging variant that stops 60% of the warm-started environments partway along the home-to-hover path; episode terminates if the knife falls off the table.
pen_holder
Finger targets limited to 0–20 mm with a first-order filter (coefficient 0.9), since a single noisy finger action dropped an in-hand pen; measured 7.5 mm hold width (the collider is far thinner than the visual barrel); warm-start hover target above the focused pen, with optional in-hand and re-yawed pen starts.
tool_packing
A drawer-pull ladder (reach handle, grasp handle, pull-out ramp with a bonus once 7 cm of tray is exposed) placed before the per-item pick-and-drop ladder; the stapler as the single active item; warm-start hover target in front of the drawer handle, or above the item once its drawer is open; a curriculum that resets half of the environments with the assigned drawer already open.
Appendix
Table 17: Per-task environment changes in the expert-tuned condition. Measured quantities were obtained by probing the simulator. Curricula apply to training only; grading always starts from the home pose.
Task
Horizon
Steps
Ent.
Finger kp
Warm start
Jitter (cm)
bulb
14
32
0.006
8,000
50%, 90
3
nut_thread
14
32
0.006
8,000
50%, 90
3
slice
20
64
0.006
2,000
60%, 100
2
pen_holder
14
32
0.006
2,000
50%, 90
3
tool_packing
24
32
0.006
8,000
50%, 90
3
Appendix
Table 18: Fine-tuned training hyperparameters in the expert-tuned condition. PPO settings not listed keep the values in the text. The warm-start curriculum pre-rolls the given fraction of environments to the hover target for the given number of control steps, with lateral jitter, before each episode.
Target
Similar source
Shared skill
Dissimilar source
bulb
nut_thread
screwing along a thread
tool_packing
allen_bolt
nut_thread
threading/ratcheting a screw
pen_holder
pc_motherboard
allen_bolt
hex-key bolt ratcheting
spatula
pc_ram
pc_gpu
seating a card in a slot
bulb
tool_packing
pen_holder
stowing objects into recesses
nut_thread
spatula
allen_bolt
wielding a grasped tool
pc_gpu
Appendix
Table 19: Cross-task transfer pairs. Similar sources share the target’s dominant contact skill; dissimilar sources have a different dominant skill.
Experiment
Setting
Score
Success
Tokens per run (M)
Unc. input
Cached input
Output
Cross-task
similar-task hint
0.60±0.05
31%
1.9
54
0.13
dissimilar-task hint
0.51±0.05
17%
1.9
86
0.15
no hint
0.52±0.05
11%
2.3
84
0.15
Cross-embodiment
Kinova Gen3 + hint
0.74±0.05
56%
0.8
56
0.10
xArm7 + hint
0.66±0.06
36%
1.3
55
0.10
Appendix
Table 20: Transfer learning results. Final rubric score (mean ± SEM over 36 runs), success rate (fraction of runs ending at rubric score 1.0 ), and mean token use per run in millions. Both no-hint rows use the same from-scratch Franka runs on the six target tasks.
Figure 17: Cross-task generalization by target task. Rubric score versus agent wall-clock time (mean ± SEM over 2 models × 3 seeds) for the similar-task hint, dissimilar-task hint, and no-hint settings.
Figure 18: Cross-embodiment generalization by task. Rubric score versus agent wall-clock time (mean ± SEM over 2 models × 3 seeds) for the two target embodiments given the Franka reference solution as a hint. From-scratch Franka runs are shown as a reference for task difficulty.
Number of demonstrations
Task
10
50
100
200
400
Success rate (%)
coffee
3.10±3.10
6.20±4.30
25.00±7.70
31.00±8.20
34.00±8.40
pc_gpu
0.00±0.00
41.00±12.00
34.00±21.00
69.00±15.00
81.00±11.00
clear_organic_objects
16.00±24.00
34.00±28.00
41.00±6.00
41.00±12.00
50.00±18.00
pen_holder
3.00±3.00
49.00±9.00
34.00±8.00
50.00±9.00
59.00±9.00
Appendix
Table 21: VLA finetuning results. Success rates (%) and task scores at five training-set sizes. Revised success criteria for selected tasks are listed in Table 22 .
Task
Revised success criterion
Coffee preparation
Press START with the bay cover closed, the capsule seated, and the mug positioned under the spout.
Clearing organic objects
Complete the placement of one organic object.
Pen holding
Insert one pen into the holder.
Banana slicing
Cut the banana into at least three pieces.
Appendix
Table 22: Task-specific evaluation criteria. For some tasks, we relax the original success conditions as follows.
Figure 19: Illustration of the real setup versus the constructed sim setup.
Figure 20: RL training curves.
Figure 21: Failure mode 1a: Opus 5 hallucinates progress on Syringe task.
Figure 22: Failure mode 1b: Opus 5 fails to notice pathological states on the pen_holder task.
Figure 23: Failure mode 2: Opus 4.8 over-commits to a failing plan on the Syringe task.
Figure 24: Failure mode 3: GPT-5.6-Sol hacks the scorer on the Syringe task.
Coding agents are beginning to move beyond purely digital tasks to tackle physical-world challenges, particularly in robotics. Existing robotics benchmarks, however, primarily focus on the performance of individual artifacts, such as policies or controllers, offering limited coverage of coding agents' broader engineering capabilities. Real-world robotics extends beyond control: agents must build, integrate, diagnose, and improve heterogeneous artifacts under resource constraints and reason from multimodal feedback. To evaluate these broader capabilities, we introduce RLE-Bench, a benchmark of robot-learning tasks spanning four representative robotics development workflows: interactive control, policy learning, perception and estimation, and mechanical design. We use diverse task-specific metrics to evaluate the artifacts submitted by the coding agents, from the success rate the agents achieved to the policy agents trained, the harness agent built, and the mechanical structures the agent designed. We aggregate these metrics into an overall RLE Index and report workflow-specific capability profiles, enabling systematic comparison of coding agents' capabilities across multiple capability dimensions. Beyond performance ranks, we also conduct in-depth case studies examining agent behavior on representative tasks, highlighting both current capabilities and limitations, and pointing to the opportunities robotics tasks have to offer for future agent training.
Haitong Ma, Chenxiao Gao, Rushi Qiang +2
Harvard University · Georgia Institute of Technology
Achieving dexterous robotic manipulation in the real world heavily relies on human supervision and algorithm engineering, which becomes a central bottleneck in the pursuit of general physical intelligence. Although emerging coding agents can generate code to automate algorithm search, their successes remain largely confined in digital environments. We conjecture that the missing abstraction to automate robotics research is a repeatable feedback loop for real-world policy improvement: reset the scene, execute a policy, verify the outcome, and refine the next iteration. To bridge this gap, we introduce ENPIRE, a harness framework for coding agents that instantiates this physical feedback routine with four core modules: an Environment module (EN) for automatic reset and verification, a Policy Improvement module (PI) that launches policy refinement, a Rollout module (R) to evaluate policies with one or multiple physical robots operating in parallel, and an Evolution module (E) in which coding agents analyze logs, consult literature, improve training infrastructure and algorithm code to address failure modes. This closed-loop system transforms real-world manipulation learning into a controllable optimization procedure, minimizing human effort while allowing fair ablations across training recipe and agent variants. Powered by ENPIRE, frontier coding agents can autonomously train a policy to achieve a 99% success rate on challenging, dexterous manipulation tasks, such as organizing a pin box, fastening a zip tie, and tool use, a process that further accelerates when we dispatch an agent team on a robot fleet. Our results suggest a practical and scalable path toward deploying coding agents to autonomously advancing robotics in the physical world.
Coding agents can construct executable programs and revise them through feedback, offering a foundation for embodied control without target-task fine-tuning. Through a programmatic robot interface, the agent observes the scene, executes control programs, and adapts them to the resulting physical state, serving as the closed-loop policy. However, this within-episode adaptation does not automatically preserve experience across scenes. Our trajectory analysis shows that agents spend interaction and inference budgets establishing effective operations and identifying their preconditions, while fresh agents cannot reuse these discoveries. This motivates EmbodiedEvo, a self-evolving coding agent that distills physical feedback into reusable code Skills and diagnostic Lessons. Experience is acquired on one Playground scene per task, then frozen for independent agents in unseen scenes, without updating model weights. On RoboDojo-Sim, EmbodiedEvo achieves an average score of 44.48 and a success rate of 37.82%, surpassing all officially verified results. Paired tests show the largest gains on memory-dependent tasks and successful reuse of coordinated procedures, while fine contact control remains difficult. Experience from stronger models also improves weaker executors: Astra's experience raises Sol and Luna's success rates by 8.86 and 5.50 percentage points over no experience. Compared with raw interaction history, Skills and Lessons achieve comparable task performance at 53.0% lower execution cost.