LACE-CRAFT: Robot Co-Design with Actor Inheritance and Blackboard Collaboration

Yuhan Wen; Jiawei Wang; Qixuan Zhang; Yusen Qin; Lan Xu

cs.ROOct 7, 2026

LACE-CRAFT: Robot Co-Design with Actor Inheritance and Blackboard Collaboration

Authors: Yuhan Wen, Jiawei Wang, Qixuan Zhang, Yusen Qin, Lan Xu

Organizations: Deemos Corporation · ShanghaiTech University · University of Cambridge · D-Robotics

Abstract

Robot co-design couples morphology search with policy learning, yet training every new design from scratch discards acquired control experience. We present LACE-CRAFT, which compares continued learning on the current robot with policy adaptation to new morphology-reward pairs. LACE resumes the incumbent's full learning state and initializes compatible challengers with its actor parameters and observation statistics. A fixed task metric selects among both branches and the frozen incumbent. CRAFT coordinates Feedback, Morphology, Reward, and Integration roles through shared experimental records and behavioral replays to generate and cross-review paired proposals. A generative extension converts generated meshes into editable articulated models with configured joints, actuator interfaces, and consistently updated simulation assets. Across five locomotion benchmarks, mean scores over three evaluation seeds are 6.4-91.9% higher than D2C. Both methods train 30 new morphology-reward pairs over five rounds; LACE additionally uses four continuation training units. Five-task ablations examine policy inheritance and replay-derived feedback. A fabricated prototype demonstrates indoor walking and illustrates the geometry-to-hardware workflow.

Figures & tables

Component	Continuation	Challenger
Morphology and reward	Unchanged	New proposal
Actor parameters	Restore	Copy
Observation statistics	Restore	Copy
Value / critic networks	Restore	Initialize anew
Optimizer states	Restore	Initialize anew
SAC targets and temperature	Restore	Initialize anew

TABLE I: Training-state handling before each branch learns. Copied Actor parameters and observation statistics continue to update.

Record	Shared contents
Experiment history	Candidate identifier, selected score, and linked morphology, reward, and checkpoint files.
Current evidence	Saved replay video, trajectory and task measurements, and the selected-checkpoint reference.
Role	Output
Feedback	Failure hypotheses.
Morphology	Body edits and reward cross-review.
Reward	Reward code and body cross-review.

TABLE II: Shared CRAFT evidence and role-specific outputs.

Fig. 2: An input image and user requirement define editable morphology $m$ through Rodin reconstruction and BANG decomposition. The four views show complementary components, not sequential steps. The base link incorporates an SO-101 mount, while other arm components reuse the SO-101 design. CRAFT supplies morphology edits and training rewards. LACE receives the compiled model and selects trained candidates using a task metric distinct from training reward. Replays and scores return to CRAFT. Simulation evaluates walking with fixed arm postures, not manipulation. Physical walking excludes the arm. The mounted-arm photograph documents assembly only.

Fig. 3: Three-seed evaluation means with

\pm 1

sample-standard-deviation error bars. Means and deviations are divided by the task’s fixed D2C mean; denominator uncertainty is not propagated.

Task	Default	D2C	LACE–CRAFT
Ant	27,362.80 $\pm 621.94$	33,044.51 $\pm 477.05$	63,396.82 $\pm 656.01$
HalfCheetah	18,343.94 $\pm 201.20$	20,072.15 $\pm 18.82$	27,455.51 $\pm 22.39$
Hopper	4,377.53 $\pm 9.80$	6,195.22 $\pm 2.29$	7,125.95 $\pm 11.94$
Swimmer	897.83 $\pm 8.19$	6,637.06 $\pm 10.67$	9,733.43 $\pm 68.55$
Walker2D	10,114.19 $\pm 10.37$	17,362.33 $\pm 221.69$	18,466.68 $\pm 29.89$

TABLE III: Final-policy evaluation scores: mean with

\pm

sample standard deviation below, over three evaluation seeds (128 episodes each).

Fig. 4: Three-seed evaluation means with

\pm 1

sample-standard-deviation error bars, divided by the task’s fixed full-method mean. Denominator uncertainty is not propagated.

Task	Full	No Actor	No replay
Ant	63,396.82 $\pm 656.01$	54,806.01 $\pm 432.08$	50,609.18 $\pm 838.12$
HalfCheetah	27,455.51 $\pm 22.39$	25,147.09 $\pm 116.61$	25,272.34 $\pm 25.86$
Hopper	7,125.95 $\pm 11.94$	6,600.18 $\pm 9.61$	6,693.74 $\pm 2.87$
Swimmer	9,733.43 $\pm 68.55$	4,278.96 $\pm 23.92$	4,832.12 $\pm 2.60$
Walker2D	18,466.68 $\pm 29.89$	10,919.33 $\pm 133.34$	13,171.87 $\pm 232.93$

TABLE IV: Final-policy ablations: mean with

\pm

sample standard deviation below, over three evaluation seeds (128 episodes each).

Fig. 5: Hardware-adapted printable parts and assembly with the SO-101 arm.

Fig. 6: Indoor walking with the arm removed. Frames are from one video; timestamps are in seconds.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Continuation	Challenger
Morphology and reward	Unchanged	Proposed pair
Actor parameters	Restore	Copy
Observation statistics	Restore	Copy
Value / critic parameters	Restore	Initialize anew
Optimizer states	Restore	Initialize anew
SAC targets and entropy state	Restore	Initialize anew

Appendix

TABLE S1: Checkpoint handling for the current LACE protocol.

Record	Stored content and use
Selected robot	Candidate and checkpoint identifiers, morphology and reward bindings, score, and full-state checkpoint.
Selection history $\mathcal{A}_{k}$	Compared candidates, scores, and round-specific selection flags; the bounded view enters the next discussion.
Evidence $\mathcal{E}_{k}$	Selected-checkpoint rollout summary and references to replay and visual manifests.
Role archive $\mathcal{H}_{k}$	Requests, responses, drafts, revisions, integration rationales, validation errors, and API records; retained separately from future role inputs.

Appendix

TABLE S2: Persistent records and their use in CRAFT.

Setting	Ant	HC	Swimmer
Learning rate	$3\times 10^{-4}$	$3\times 10^{-4}$	$3\times 10^{-4}$
Discount	0.97	0.95	0.97
Entropy cost	0.01	0.001	0.001
Reward scaling	25	3	5
Training environments	256	128	128
Batch size	512	512	128

Appendix

TABLE S3: PPO configuration. HC denotes HalfCheetah.

Setting	Hopper	Walker2D
Learning rate	$6\times 10^{-4}$	$6\times 10^{-4}$
Discount	0.997	0.997
Reward scaling	30	5
Training environments	12	12
Batch size	48	48
Gradient updates per step	3	3

Appendix

TABLE S4: SAC configuration.

Task	RL	Requested	Actual	Horizon
Ant	PPO	8,000,000	8,007,680	450
HalfCheetah	PPO	5,000,000	5,038,080	400
Hopper	SAC	400,000	400,008	800
Swimmer	PPO	350,000	350,208	600
Walker2D	SAC	500,000	500,004	1,000

Appendix

TABLE S5: Task-specific learning configurations. Actual transition counts include batch alignment.

Index	Interval	Index	Interval
1	$[-.90,-.20]$	13	$[-.35,.50]$
2	$[.20,.90]$	14	$[-.45,-.04]$
3	$[.55,1.35]$	15	$[-.50,.35]$
4	$[.04,.45]$	16	$[-.35,-.02]$
5	$[-.50,.30]$	17	$[.015,.100]$
6	$[-.45,-.04]$	18	$[.010,.090]$

Appendix

TABLE S6: HalfCheetah morphology-vector intervals.

Task	Coordinates	Challenger	Continuation
Ant	$(x,y,z)$	Copy	Full state
HalfCheetah	$(x,y,z)$	Copy	Full state
Hopper	$(x,y,z)$	Copy	Full state
Swimmer	$(x,y)$	Copy	Full state
Walker2D	$(x,y,z)$	Copy	Full state

Appendix

TABLE S7: Initialization of the reported five-task searches.

Fig. S1: Ant candidate scores in rounds 2–4 and the five-round selection path. Stars mark winners. Inc.: incumbent; Cont.: continuation; C1–C6: challengers ordered by identifier within each round.

Search	RL seed	Selected score
1	0	57,560.79
2	0	55,259.83
3	0	64,252.91

Appendix

TABLE S8: Additional Full searches on Ant; 30 generation units per search. Scores are historical selection scores.

Fig. S2: Exploratory search-level variation on Ant versus cumulative generation units. Full shows the mean and sample standard deviation of three searches at RL seed 0; other methods have one search each.

Round	D2C	LACE–CRAFT	Gain
1	11,023.11	12,435.48	$+12.8\%$
2	14,422.04	16,488.11	$+14.3\%$
3	14,902.87	16,742.91	$+12.3\%$
4	13,902.85	15,088.00	$+8.5\%$
5	13,186.34	16,394.60	$+24.3\%$

Appendix

TABLE S9: Walker2D: best new design in each round.

Task	Variant	Score	Units	Cont.
Ant	Full	64,252.91	34	4
	No Actor	56,901.25	34	4
	No replay	51,878.46	34	4
HalfCheetah	Full	27,458.55	34	4
	No Actor	25,263.65	34	4
	No replay	25,277.06	34	4

Appendix

TABLE S10: Historical search-selection scores and completed training budgets. These are not the post-selection three-seed means in the main text. Units include continuations.

Task	Variant	Round 1	Round 2	Round 3	Round 4	Round 5
Ant	No Actor	34,631.61	50,914.84	50,914.84	56,901.25	56,901.25
	No replay	29,049.02	42,973.31	48,472.22	51,878.46	51,878.46
HalfCheetah	No Actor	25,263.65	25,263.65	25,263.65	25,263.65	25,263.65
	No replay	20,733.75	23,762.41	24,715.28	25,277.06	25,277.06
Hopper	No Actor	5,212.28	6,595.88	6,595.88	6,595.88	6,595.88
	No replay	6,573.45	6,573.45	6,573.45	6,573.45	6,693.28

Appendix

TABLE S11: Selected scores after each ablation round. Repeated values indicate retention of the best available robot.

Stage	Output
Shape generation	Static mesh conditioned on the morphology reference.
Decomposition and alignment	Separate link meshes $g_{i}$ with local frames and aligned connections.
Joint allocation	Parent–child graph, joint origins, types, axes, and position limits.
Actuator fitting	Motor positions and orientations, cavity geometry, and mating surfaces in $\Sigma$ .
Compilation	URDF, collision geometry, inertial properties, joint and actuator limits, and action mapping in $\Phi$ .

Appendix

TABLE S12: Artifacts produced by generative-morphology processing.

Fig. S3: Ant single-search progress versus cumulative generation units. Each curve shows stored selection scores from one search, not the post-selection evaluation means.

Explore similar work

Debate2Create: Robot Co-design via Multi-Agent LLM Debate
Oct 29, 2025Kevin Qiu, Marek CyganMulti-Agent Debate Multi-Agent Large Language Model Systems
University of Warsaw · 2IDEAS NCBR · 3Nomagic
ECo-MoE: Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots
May 22, 2026Yibin Wang, Muhan Li, Zihan Guo +1Co-Evolution Robot Systems
Northwestern University, Evanston, IL, USA.
RoboCoach: World Models as Active Coaches for Compositional Robot Skills
Sep 30, 2026Jiajun Liu, Yifan Chen, Yichao Liu +7Robot Systems Robocasa
Renmin University of China · Tsinghua University · Shanghai Qizhi Institute +2

Oct 29, 2025, cs.ROJ/K move · Enter open · S save

Debate2Create: Robot Co-design via Multi-Agent LLM Debate

Kevin Qiu, Marek Cygan

University of Warsaw · 2IDEAS NCBR · 3Nomagic

Multi-Agent Debate Multi-Agent Large Language Model Systems Debate Mujoco

Open paper AskPDF arXiv

We introduce Debate2Create (D2C), a multi-agent LLM framework that formulates robot co-design as structured, iterative debate grounded in physics-based evaluation. A design agent and control agent engage in a thesis-antithesis-synthesis loop, while criterion-specific LLM judges provide multi-objective feedback to steer exploration. Across five MuJoCo locomotion benchmarks, D2C achieves the highest default-normalized score among the evaluated LLM-based and black-box baselines, with gains up to 3.2x on Ant and nearly 9x on Swimmer. Iterative debate yields 18-35% gains over compute-matched zero-shot generation, and D2C-generated rewards transfer to default morphologies in 4/5 tasks. These results suggest that structured, simulator-grounded multi-agent interaction is a useful mechanism for joint morphology-reward optimization under a fixed-topology, per-candidate-RL protocol. Project page: debate2create.github.io.

Abstract

Explore similar work

Debate2Create: Robot Co-design via Multi-Agent LLM Debate

ECo-MoE: Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots

RoboCoach: World Models as Active Coaches for Compositional Robot Skills